Re: [PATCH 0/3] KVM: x86: Fix supported VM_TYPES caps

[Sean Christopherson]

On Mon, Apr 29, 2024, Kai Huang wrote:
> On 27/04/2024 3:47 am, Sean Christopherson wrote:
> > On Fri, Apr 26, 2024, Kai Huang wrote:
> > > On Thu, 2024-04-25 at 07:30 -0700, Sean Christopherson wrote:
> > > > On Thu, Apr 25, 2024, Xiaoyao Li wrote:
> > > > > On 4/24/2024 12:53 AM, Sean Christopherson wrote:
> > > > > > Fix a goof where KVM fails to re-initialize the set of supported VM types,
> > > > > > resulting in KVM overreporting the set of supported types when a vendor
> > > > > > module is reloaded with incompatible settings.  E.g. unload kvm-intel.ko,
> > > > > > reload with ept=0, and KVM will incorrectly treat SW_PROTECTED_VM as
> > > > > > supported.
> > > > > 
> > > > > Hah, this reminds me of the bug of msrs_to_save[] and etc.
> > > > > 
> > > > >     7a5ee6edb42e ("KVM: X86: Fix initialization of MSR lists")
> > > > 
> > > > Yeah, and we had the same bug with allow_smaller_maxphyaddr
> > > > 
> > > >    88213da23514 ("kvm: x86: disable the narrow guest module parameter on unload")
> > > > 
> > > > If the side effects of linking kvm.ko into kvm-{amd,intel}.ko weren't so painful
> > > > for userspace,
> > > > 
> > > 
> > > Do we have any real side effects for _userspace_ here?
> > 
> > kvm.ko ceasing to exist, and "everything" being tied to the vendor module is the
> > big problem.  E.g. params from the kernel command line for kvm.??? will become
> > ineffective, etc.  Some of that can be handled in the kernel, e.g. KVM can create
> > a sysfs symlink so that the accesses through sysfs continue to work, but AFAIK
> > params don't supporting such aliasing/links.
> > 
> > I don't think there are any deal breakers, but I don't expect it to Just Work either.
> 
> Perhaps we can make the kvm.ko as a dummy module which only keeps the module
> parameters for backward compatibility?

Keeping parameters in a dummy kvm.ko would largely defeat the purpose of linking
everything into vendor modules, i.e. would make it possible for the parameters to
hold a stale value.