Re: [PATCH v3 2/2] VMX: reset the segment cache after segment initialization in vmx_vcpu_reset

[Maxim Levitsky <mlevitsk@redhat.com>]

On Fri, 2024-08-09 at 08:27 -0700, Sean Christopherson wrote:
> On Tue, Aug 06, 2024, Chao Gao wrote:
> > On Thu, Jul 25, 2024 at 01:52:32PM -0400, Maxim Levitsky wrote:
> > > Fix this by moving the vmx_segment_cache_clear() call to be after the
> > > segments are initialized.
> > > 
> > > Note that this still doesn't fix the issue of kvm_arch_vcpu_in_kernel
> > > getting stale data during the segment setup, and that issue will
> > > be addressed later.
> > > 
> > > Signed-off-by: Maxim Levitsky <mlevitsk@redhat.com>
> > 
> > Do you need a Fixes tag and/or Cc: stable?
> 
> Heh, it's an old one
> 
>   Fixes: 2fb92db1ec08 ("KVM: VMX: Cache vmcs segment fields")
> 
> > > ---
> > > arch/x86/kvm/vmx/vmx.c | 6 +++---
> > > 1 file changed, 3 insertions(+), 3 deletions(-)
> > > 
> > > diff --git a/arch/x86/kvm/vmx/vmx.c b/arch/x86/kvm/vmx/vmx.c
> > > index fa9f307d9b18..d43bb755e15c 100644
> > > --- a/arch/x86/kvm/vmx/vmx.c
> > > +++ b/arch/x86/kvm/vmx/vmx.c
> > > @@ -4870,9 +4870,6 @@ void vmx_vcpu_reset(struct kvm_vcpu *vcpu, bool init_event)
> > > 	vmx->hv_deadline_tsc = -1;
> > > 	kvm_set_cr8(vcpu, 0);
> > > 
> > > -	vmx_segment_cache_clear(vmx);
> > > -	kvm_register_mark_available(vcpu, VCPU_EXREG_SEGMENTS);
> > > -
> > > 	seg_setup(VCPU_SREG_CS);
> > > 	vmcs_write16(GUEST_CS_SELECTOR, 0xf000);
> > > 	vmcs_writel(GUEST_CS_BASE, 0xffff0000ul);
> > > @@ -4899,6 +4896,9 @@ void vmx_vcpu_reset(struct kvm_vcpu *vcpu, bool init_event)
> > > 	vmcs_writel(GUEST_IDTR_BASE, 0);
> > > 	vmcs_write32(GUEST_IDTR_LIMIT, 0xffff);
> > > 
> > > +	vmx_segment_cache_clear(vmx);
> > > +	kvm_register_mark_available(vcpu, VCPU_EXREG_SEGMENTS);
> > 
> > vmx_segment_cache_clear() is called in a few other sites. I think at least the
> > call in __vmx_set_segment() should be fixed, because QEMU may read SS.AR right
> > after a write to it. if the write was preempted after the cache was cleared but
> > before the new value being written into VMCS, QEMU would find that SS.AR held a
> > stale value.
> 
> Ya, I thought the plan was to go for a more complete fix[*]?  This change isn't
> wrong, but it's obviously incomplete, and will be unnecessary if the preemption
> issue is resolved.

Hi,

I was thinking to keep it simple, since the issue is mostly theoretical after this fix,
but I'll give this another try.

Best regards,
	Maxim Levitsky

> 
> [*] https://lore.kernel.org/all/f183d215c903d4d1e85bf89e9d8b57dd6ce5c175.camel@redhat.com
>