Re: [PATCH 1/7] KVM: VMX: remove regs argument of __vmx_vcpu_run

From: Josh Poimboeuf
Date: Tue Nov 01 2022 - 13:34:12 EST


On Mon, Oct 31, 2022 at 05:37:46PM +0000, Sean Christopherson wrote:
> On Fri, Oct 28, 2022, Paolo Bonzini wrote:
> > Registers are reachable through vcpu_vmx, no need to pass
> > a separate pointer to the regs[] array.
> >
> > No functional change intended.
> >
> > Signed-off-by: Paolo Bonzini <pbonzini@xxxxxxxxxx>
> > ---
> > arch/x86/kernel/asm-offsets.c | 1 +
> > arch/x86/kvm/vmx/nested.c | 3 +-
> > arch/x86/kvm/vmx/vmenter.S | 58 +++++++++++++++--------------------
> > arch/x86/kvm/vmx/vmx.c | 3 +-
> > arch/x86/kvm/vmx/vmx.h | 3 +-
> > 5 files changed, 29 insertions(+), 39 deletions(-)
> >
> > diff --git a/arch/x86/kernel/asm-offsets.c b/arch/x86/kernel/asm-offsets.c
> > index cb50589a7102..90da275ad223 100644
> > --- a/arch/x86/kernel/asm-offsets.c
> > +++ b/arch/x86/kernel/asm-offsets.c
> > @@ -111,6 +111,7 @@ static void __used common(void)
> >
> > if (IS_ENABLED(CONFIG_KVM_INTEL)) {
> > BLANK();
> > + OFFSET(VMX_vcpu_arch_regs, vcpu_vmx, vcpu.arch.regs);
>
> Is there an asm-offsets-like solution that doesn't require exposing vcpu_vmx
> outside of KVM? We (Google) want to explore loading multiple instances of KVM,
> i.e. loading multiple versions of kvm.ko at the same time, to allow intra-host
> migration between versions of KVM to upgrade/rollback KVM without changing the
> kernel (RFC coming soon-ish). IIRC, asm-offsets is the only place where I haven't
> been able to figure out a simple way to avoid exposing KVM's internal structures
> outside of KVM (so that the structures can change across KVM instances without
> breaking kernel code).

Is that really a problem? Would it even make sense for non-KVM kernel
code to use 'vcpu_vmx' anyway? It already seems to be private.
asm-offsets.c has to "cheat" to get access to it by including
"../kvm/vmx/vmx.h".

So the only concern is exposing the asm offsets, right? But it seems
highly unlikely any non-KVM code would be using those either.

And, that would be a bug anyway: module code is subject to change and
could always get recompiled. The kernel proper shouldn't be making any
assumptions about the layouts of module-owned structs.

--
Josh