On Fri, Apr 26, 2024, Michael Roth wrote:
> On Fri, Apr 26, 2024 at 05:10:10PM -0700, Sean Christopherson wrote:
> > e.g. put the cert in a directory along with a lock. Actually, IIUC, there doesn't
> > even need to be a separate lock file. I know very little about userspace programming,
> > but common sense and a quick search tells me that file locks are a solved problem.
> >
> > E.g. it took me ~5 minutes of Googling to come up with this, which AFAICT does
> > exactly what you want.
> >
> > touch ~/vlek.cert
> > (
> > flock -e 200
> > echo "Locked the cert, sleeping for 10 seconds"
> > sleep 10
> > echo "Igor, it's alive!!!!!!"
> > ) 200< vlek.cert
> >
> > touch ~/vlek.cert
> > (
> > flock -s 201
> > echo "Got me a shared lock, no updates for you!"
> > ) 201< vlek.cert
> >
>
> Hmm... I did completely miss this option. But I think there are still some
> issues here. IIUC you're suggesting (for example):
>
> "Management":
> a) writelock vlek.cert
> b) perform SNP_LOAD_VLEK and update vlek.cert contents
> c) unlock vlek.cert
>
> "QEMU":
> a) readlock vlek.cert
> b) copy cert into guest buffer
> c) unlock vlek.cert
>
> The issue is that after "QEMU" unlocks and return the cert to KVM we'll
> have:
>
> "KVM"
> a) return from EXT_GUEST_REQ exit to userspace
> b) issue the attestation report to firmware
> c) return the attestation report and cert to the guest
>
> Between a) and b), "Management" can complete another entire update, but
> the cert that it passes back to the guest will be stale relative to the
> key used to sign the attestation report.
I was thinking userspace would hold the lock across SEV_CMD_SNP_GUEST_REQUEST.
QEMU:
a) readlock vlek.cert
b) copy cert into guest buffer
c) set kvm_run->immediate_exit
d) invoke KVM_RUN
e) KVM sends SEV_CMD_SNP_GUEST_REQUEST to PSP
f) KVM exits to userspace with -EINTR
g) unlock vlek.cert
h) invoke KVM_RUN (resume the guest)
> If we need to take more time to explore other options it's not
> absolutely necessary to have the kernel solve this now. But every userspace
> will need to solve it in some way so it seemed like it might be nice to
> have a simple reference implementation to start with.
Shoving something into the kernel is not a "reference implementation", especially
not when it impacts the ABI of multiple subsystems.
