On Wed, Mar 29, 2023 at 11:17:31PM +0000,
"Huang, Kai" <kai.huang@xxxxxxxxx> wrote:
> On Wed, 2023-03-29 at 14:56 -0700, Isaku Yamahata wrote:
> > On Wed, Mar 29, 2023 at 01:13:45AM +0000,
> > "Huang, Kai" <kai.huang@xxxxxxxxx> wrote:
> >
> > > >
> > > > > >
> > > > > > + /*
> > > > > > + * TDX requires those methods to enable VMXON by
> > > > > > + * kvm_hardware_enable/disable_all()
> > > > > > + */
> > > > > > + static_call_update(kvm_x86_check_processor_compatibility,
> > > > > > + ops->runtime_ops->check_processor_compatibility);
> > > > > > + static_call_update(kvm_x86_hardware_enable,
> > > > > > + ops->runtime_ops->hardware_enable);
> > > > > > + static_call_update(kvm_x86_hardware_disable,
> > > > > > + ops->runtime_ops->hardware_disable);
> > > > > > r = ops->hardware_setup();
> > > > > > if (r != 0)
> > > > > > goto out_mmu_exit;
> > > > >
> > > > > Hmm.. I think this is ugly. Perhaps we should never do any
> > > > > static_call(kvm_x86_xxx)() in hardware_setup(), because hardware_setup() is
> > > > > called before kvm_ops_update() and may update vendor's kvm_x86_ops.
> > > > >
> > > > > So probably use hardware_enable_all() in hardware_setup() is a bad idea.
> > > > >
> > > > > I think we have below options on how to handle:
> > > > >
> > > > > 1) Use VMX's kvm_x86_ops directly in tdx_hardware_setup(). For instance,
> > > > > something like below:
> > > > >
> > > > > int __init tdx_hardware_setup(struct kvm_x86_ops *x86_ops)
> > > > > {
> > > > > ...
> > > > >
> > > > > cpus_read_lock();
> > > > > r = on_each_cpu(vt_x86_ops.hardware_enable, ...);
> > > > > if (!r)
> > > > > r = tdx_module_setup();
> > > > > on_each_cpu(vt_x86_ops.hardware_disable, ...);
> > > > > cpus_read_unlock();
> > > > >
> > > > > ...
> > > > > }
> > > > >
> > > > > But this doesn't clean up nicely when there's some particular cpus fail to do
> > > > > hardware_enable(). To clean up nicely, we do need additional things similar to
> > > > > the hardware_enable_all() code path: a per-cpu variable or a cpumask_t + a
> > > > > wrapper of vt_x86_ops->hardware_enable() to track which cpus have done
> > > > > hardware_enable() successfully.
> > > > >
> > > > > 2) Move those static_call_update() into tdx_hardware_setup() so they are TDX
> > > > > code self-contained. But this would require exposing kvm_x86_ops as symbol,
> > > > > which isn't nice either.
> > > > >
> > > > > 3) Introduce another kvm_x86_init_ops->hardware_post_setup(), which is called
> > > > > after kvm_ops_update().
> > > > >
> > > > > Personally, I think 3) perhaps is the most elegant one, but not sure whether
> > > > > Sean/Paolo has any opinion.
> > > >
> > > > I think we can simply update the ops before calling hardware_enable() and
> > > > clean up ops on failure.
> > > >
> > > >
> > >
> > > This doesn't work because hardware_setup() may update vendor's kvm_x86_ops.
> > >
> > > If you do kvm_ops_update() before hardware_setup(), you need to manually update
> > > those updated (in hardware_setup()) callbacks again after.
> >
> > We can call kvm_ops_update() twice before and after hardware_setup().
> >
>
> Personally I think it's too ugly.
So you prefer the option 3 to calling kvm_ops_update() twice. Okay, I'll update
the patch.
--
Isaku Yamahata <isaku.yamahata@xxxxxxxxx>
