Re: [PATCH] KVM: x86: Only print persistent reasons for kvm disabled once

["Radim Krčmář" <rkrcmar@redhat.com>]

2019-08-27 08:27+0200, Vitaly Kuznetsov:
> Tony Luck <tony.luck@intel.com> writes:
> 
> > When I boot my server I'm treated to a console log with:
> >
> > [   40.520510] kvm: disabled by bios
> > [   40.551234] kvm: disabled by bios
> > [   40.607987] kvm: disabled by bios
> > [   40.659701] kvm: disabled by bios
> > [   40.691224] kvm: disabled by bios
> > [   40.718786] kvm: disabled by bios
> > [   40.750122] kvm: disabled by bios
> > [   40.797170] kvm: disabled by bios
> > [   40.828408] kvm: disabled by bios
> >
> >  ... many, many more lines, one for every logical CPU
> 
> (If I didn't miss anything) we have the following code:
> 
> __init vmx_init()
>         kvm_init();
>             kvm_arch_init()
> 
> and we bail on first error so there should be only 1 message per module
> load attempt. The question I have is who (and why) is trying to load
> kvm-intel (or kvm-amd which is not any different) for each CPU? Is it
> udev? Can this be changed?

I agree that this is a highly suspicious behavior.

It would be really helpful if we found out what is causing it.
So far, this patch seems to be working around a userspace bug.

> In particular, I'm worried about eVMCS enablement in vmx_init(), we will
> also get a bunch of "KVM: vmx: using Hyper-V Enlightened VMCS" messages
> if the consequent kvm_arch_init() fails.

And we can't get rid of this through the printk_once trick, because this
code lives in kvm_intel module and therefore gets unloaded on every
failure.

I am also not inclined to apply the patch as we will likely merge the
kvm and kvm_{svm,intel} modules in the future to take full advantage of
link time optimizations and this patch would stop working after that.

Thanks.