Re: [RFC PATCH v1 0/5] x86/boot, KVM: Move VMXON/VMXOFF handling from KVM to CPU lifecycle

[Sean Christopherson <seanjc@google.com>]

On Wed, Sep 17, 2025, Xin Li wrote:
> On 9/16/2025 10:54 AM, Sean Christopherson wrote:
> > On Thu, Sep 11, 2025, Arjan van de Ven wrote:
> > > Hi,
> > > > I also want to keep the code as a module, both to avoid doing VMXON unconditionally,
> > > 
> > > can you expand on what the problem is with having VMXON unconditionally enabled?
> > 
> > Unlike say EFER.SVME, VMXON fundamentally changes CPU behavior.  E.g. blocks INIT,
> > activates VMCS caches (which aren't cleared by VMXOFF on pre-SPR CPUs, and AFAIK
> > Intel hasn't even publicly committed to that behavior for SPR+), restricts allowed
> > CR0 and CR4 values, raises questions about ucode patch updates, triggers unique
> > flows in SMI/RSM, prevents Intel PT from tracing on certain CPUs, and probably a
> > few other things I'm forgetting.
> 
> Regarding Intel PT, if VMXON/VMXOFF are moved to CPU startup/shutdown, as
> Intel PT is initialized during arch_initcall() stage, entering and leaving
> VMX operation no longer happen while Intel PT is _active_, thus
> intel_pt_handle_vmx() no longer needs to "handles" VMX state transitions.

The issue isn't handling transitions, it's that some CPUs don't support Intel PT
post-VMXON:

  If bit 14 is read as 1, Intel® Processor Trace (Intel PT) can be used in VMX
  operation. If the processor supports Intel PT but does not allow it to be used
  in VMX operation, execution of VMXON clears IA32_RTIT_CTL.TraceEn (see
  “VMXON—Enter VMX Operation” in Chapter 32); any attempt to write IA32_RTIT_CTL
  while in VMX operation (including VMX root operation) causes a general-protection
  exception.

And again, unconditionally doing VMXON is a minor objection in all of this.