Re: [PATCH 0/2] Improved support for AMD SEV firmware loading

["Daniel P. Berrangé" <berrange@redhat.com>]

On Mon, Jan 17, 2022 at 06:12:50AM -0600, Brijesh Singh wrote:
> 
> On 1/17/22 1:34 AM, Dov Murik wrote:
> > [+cc Tom, Brijesh, Ashish - see SEV-related changes in this series]
> >
> >
> > On 13/01/2022 18:55, Daniel P. Berrangé wrote:
> >> The AMD SEV build of EDK2 only emits a single file, intended to be
> >>
> >> mapped readonly. There is explicitly no separate writable VARS
> >>
> >> store for persisting non-volatile firmware variables.
> >>
> >>
> >>
> >> This can be used with QEMU's traditional pflash configuration
> >>
> >> mechanism by only populating pflash0, leaving pflash1 unconfigured.
> >>
> >> Conceptually, however, it is odd to be using pflash at all when we
> >>
> >> have no intention of supporting any writable variables. The -bios
> >>
> >> option should be sufficient for any firmware that is exclusively
> >>
> >> readonly code.
> >>
> >>
> >>
> >>
> >>
> >> A second issue is that the firmware descriptor schema does not allow
> >>
> >> for describing a firmware that uses pflash, without any associated
> >>
> >> non-volatile storage.
> >>
> >>
> >>
> >> In docs/interop/firmware.json
> >>
> >>
> >>
> >>  'struct' : 'FirmwareMappingFlash',
> >>
> >>   'data'   : { 'executable'     : 'FirmwareFlashFile',
> >>
> >>                'nvram-template' : 'FirmwareFlashFile' } }
> >>
> >>
> >>
> >> Notice that nvram-template is mandatory, and when consuming these
> >>
> >> files libvirt will thus complain if the nvram-template field is
> >>
> >> missing.
> >>
> >>
> >>
> >> We could in theory make nvram-template optional in the schema and
> >>
> >> then update libvirt to take account of it, but this feels dubious
> >>
> >> when we have a perfectly good way of describing a firmware without
> >>
> >> NVRAM, using 'FirmwareMappingMemory' which is intended to be used
> >>
> >> with QEMU's -bios option.
> >>
> >>
> >>
> >>
> >>
> >> A third issue is in libvirt, where again the code handling the
> >>
> >> configuration of pflash supports two scenarios
> >>
> >>
> >>
> >>  - A single pflash image, which is writable
> >>
> >>  - A pair of pflash images, one writable one readonly
> >>
> >>
> >>
> >> There is no support for a single read-only pflash image in libvirt
> >>
> >> today.
> >>
> >>
> >>
> >>
> >>
> >> This all points towards the fact that we should be using -bios
> >>
> >> to load the AMD SEV firmware build of EDK.
> >>
> >>
> >>
> >> The only thing preventing us doing that is that QEMU does not
> >>
> >> initialize the SEV firmware when using -bios. That is fairly
> >>
> >> easily solved, as done in this patch series.
> >>
> >>
> >>
> >> For testing I've launched QEMU in in these scenarios
> >>
> >>
> >>
> >>   - SEV guest using -bios and boot from HD
> >>
> >>   - SEV guest using pflash and boot from HD
> >>
> >>   - SEV-ES guest using -bios and direct kernel boot
> >>
> >>   - SEV-ES guest using pflash and direct kernel boot
> >>
> >>
> >>
> >> In all these cases I was able to validate the reported SEV
> >>
> >> guest measurement.
> >>
> >>
> >
> > I'm having trouble testing this series (applied on top of master commit 69353c332c):
> > it hangs with -bios but works OK with pflash:
> >
> > Here's with -bios:
> >
> > $ sudo /home/dmurik/git/qemu/build/qemu-system-x86_64 -enable-kvm \
> >        -cpu host -machine q35 -smp 4 -m 2G \
> >        -machine confidential-guest-support=sev0 \
> >        -object sev-guest,id=sev0,cbitpos=47,reduced-phys-bits=1,policy=0x0 \
> >        -bios /home/dmurik/git/edk2/Build/AmdSev/DEBUG_GCC5/FV/OVMF.fd \
> >        -nographic \
> >        -global isa-debugcon.iobase=0x402 -debugcon file:ovmf-1.log \
> >        -monitor pty -trace 'enable=kvm_sev_*'
> >
> > char device redirected to /dev/pts/14 (label compat_monitor0)
> > kvm_sev_init
> > kvm_sev_launch_start policy 0x0 session (nil) pdh (nil)
> > kvm_sev_change_state uninit -> launch-update
> > kvm_sev_launch_update_data addr 0x7f42e9bff010 len 0x400000
> > kvm_sev_change_state launch-update -> launch-secret
> > kvm_sev_launch_measurement data PF6n7+Vujx5sW8PC6iMRtHXfpXdJ4osbcfYvoknu7gg4ypMqs727NTzG86Ft8Llu
> > kvm_sev_launch_finish
> > kvm_sev_change_state launch-secret -> running
> >
> >
> > Here it hangs. The ovmf-1.log file is empty.
> >
> > Notice that kvm_sev_launch_update_data is called, so the new
> > -bios behaviour is triggered correctly.  But the guest doesn't
> > start running.
> 
> I have not looked at the patch detail yet but address looks wrong, it
> looks like the hva 0x7f42e9bff010 end of the ROM. We need to encrypt the
> entire ROM to boot, so I was hoping that hva will be 2MB aligned or a
> page-aligned. You can enable the KVM trace to see if we are able to
> enter and execute anything from guest.

On further investigation I think the use of -bios is probably
doomed. The way QEMU implements -bios is to load the firmware
as a ROM, then later when a CPU reset is triggered at the end
of machine construction, the ROM is copied into the RAM area.

For SEV we need the encrypted firmware in the RAM area, but
my patch here is applying the encryption in the ROM. QEMU
copies this encrypted code across in the CPU reset handler
but this can't work because IIUC that invalidates the
initialization vector preventing correct decryption.

We can't delay the encryption step until after the ROM
copying because by that time the SEV initialization sequence
has gone too far along.

I did a gross hack which disabled use of the ROM and ROM
copying and directly loaded the firmware into the RAM area
and that successfully ran, but it is too much of a hack to
consider proposing as a real patch.

I'm going to go back to just looking at how we can make
support for pflash-without-NVRAM a first class citizen
in QEMU firmware descriptors / libvirt, since that is
known to work without hacks.

So ignore this flawed series henceforth.


Regards,
Daniel
-- 
|: https://berrange.com      -o-    https://www.flickr.com/photos/dberrange :|
|: https://libvirt.org         -o-            https://fstop138.berrange.com :|
|: https://entangle-photo.org    -o-    https://www.instagram.com/dberrange :|