On Wed, 2024-04-24 at 08:33 -0700, Sean Christopherson wrote:
> On Wed, Apr 24, 2024, Kai Huang wrote:
> > On Tue, 2024-04-23 at 09:53 -0700, Sean Christopherson wrote:
> > > Zero out all of kvm_caps when loading a new vendor module to ensure that
> > > KVM can't inadvertently rely on global initialization of a field, and add
> > > a comment above the definition of kvm_caps to call out that all fields
> > > needs to be explicitly computed during vendor module load.
> > >
> > > Signed-off-by: Sean Christopherson <seanjc@xxxxxxxxxx>
> > > ---
> > > arch/x86/kvm/x86.c | 7 +++++++
> > > 1 file changed, 7 insertions(+)
> > >
> > > diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
> > > index 44ce187bad89..8f3979d5fc80 100644
> > > --- a/arch/x86/kvm/x86.c
> > > +++ b/arch/x86/kvm/x86.c
> > > @@ -92,6 +92,11 @@
> > > #define MAX_IO_MSRS 256
> > > #define KVM_MAX_MCE_BANKS 32
> > >
> > > +/*
> > > + * Note, kvm_caps fields should *never* have default values, all fields must be
> > > + * recomputed from scratch during vendor module load, e.g. to account for a
> > > + * vendor module being reloaded with different module parameters.
> > > + */
> > > struct kvm_caps kvm_caps __read_mostly;
> > > EXPORT_SYMBOL_GPL(kvm_caps);
> > >
> > > @@ -9755,6 +9760,8 @@ int kvm_x86_vendor_init(struct kvm_x86_init_ops *ops)
> > > return -EIO;
> > > }
> > >
> > > + memset(&kvm_caps, 0, sizeof(kvm_caps));
> > > +
> > > x86_emulator_cache = kvm_alloc_emulator_cache();
> > > if (!x86_emulator_cache) {
> > > pr_err("failed to allocate cache for x86 emulator\n");
> >
> > Why do the memset() here particularly?
>
> So that it happens as early as possible, e.g. in case kvm_mmu_vendor_module_init()
> or some other function comes along and modifies kvm_caps.
>
> > Isn't it better to put ...
> >
> > memset(&kvm_caps, 0, sizeof(kvm_caps));
> > kvm_caps.supported_vm_types = BIT(KVM_X86_DEFAULT_VM);
> > kvm_caps.supported_mce_cap = MCG_CTL_P | MCG_SER_P;
> >
> > ... together so it can be easily seen?
> >
> > We can even have a helper to do above to "reset kvm_caps to default
> > values" I think.
>
> Hmm, I don't think a helper is necessary, but I do agree that having all of the
> explicit initialization in one place would be better. The alternative would be
> to use |= for BIT(KVM_X86_DEFAULT_VM), and MCG_CTL_P | MCG_SER_P, but I don't
> think that would be an improvement. I'll tweak the first two patches to set the
> hardcoded caps earlier.
>
> The main reason I don't want to add a helper is that coming up with a name would
> be tricky. E.g. "kvm_reset_caps()" isn't a great fit because the caps are "reset"
> throughout module loading. "kvm_set_default_caps()" kinda fits, but they aren't
> so much that they are KVM's defaults, rather they are the caps that KVM can always
> support regardless of hardware support, e.g. supported_xcr0 isn't optional, it
> just depends on hardware.
Personally I am fine with kvm_set_default_caps(), but obviously no strong
opinion here. :-)
