Re: [RFC PATCH 1/2] kvm/x86: intercept guest changes to X86_CR4_LA57

[Sean Christopherson <sean.j.christopherson@intel.com>]

On Tue, Sep 29, 2020 at 01:32:45PM +0800, Lai Jiangshan wrote:
> On Tue, Sep 29, 2020 at 12:24 AM Sean Christopherson
> <sean.j.christopherson@intel.com> wrote:
> >
> > On Mon, Sep 28, 2020 at 04:30:46PM +0800, Lai Jiangshan wrote:
> > > From: Lai Jiangshan <laijs@linux.alibaba.com>
> > >
> > > When shadowpaping is enabled, guest should not be allowed
> > > to toggle X86_CR4_LA57. And X86_CR4_LA57 is a rarely changed
> > > bit, so we can just intercept all the attempts to toggle it
> > > no matter shadowpaping is in used or not.
> > >
> > > Fixes: fd8cb433734ee ("KVM: MMU: Expose the LA57 feature to VM.")
> > > Cc: Sean Christopherson <sean.j.christopherson@intel.com>
> > > Cc: Yu Zhang <yu.c.zhang@linux.intel.com>
> > > Cc: Paolo Bonzini <pbonzini@redhat.com>
> > > Signed-off-by: Lai Jiangshan <laijs@linux.alibaba.com>
> > > ---
> > >   No test to toggle X86_CR4_LA57 in guest since I can't access to
> > >   any CPU supports it. Maybe it is not a real problem.
> >
> 
> 
> Hello
> 
> Thanks for reviewing.
> 
> > LA57 doesn't need to be intercepted.  It can't be toggled in 64-bit mode
> > (causes a #GP), and it's ignored in 32-bit mode.  That means LA57 can only
> > take effect when 64-bit mode is enabled, at which time KVM will update its
> > MMU context accordingly.
> >
> 
> Oh, I missed that part which is so obvious that the patch
> seems impertinent.
> 
> But X86_CR4_LA57 is so fundamental that it makes me afraid to
> give it over to guests. And it is rarely changed too. At least,
> there is no better reason to give it to the guest than
> intercepting it.
> 
> There might be another reason that this patch is still needed with
> an updated changelog.
> 
> When a user (via VMM such as qemu) launches a VM with LA57 disabled
> in its cpuid on a LA57 enabled host. The hypervisor, IMO, needs to
> intercept guest's changes to X86_CR4_LA57 even when the guest is still
> in the non-paging mode. Otherwise the hypervisor failed to detective
> such combination when the guest changes paging mode later.
> 
> Anyway, maybe it is still not a real problem.

Oof, the above is a KVM bug, though in a more generic manner.  All reserved
bits should be intercepted, not just LA57.  LA57 is the only affected bit at
the moment, but proper support is needed as the follow-on patch to let the
guest toggle FSGSBASE would introduce the same bug.

Sadly, fixing this is a bit of a mess.  Well, fixing LA57 is easy, e.g. this
patch will do the trick.  But actually refreshing the CR4 guest/host mask when
the guest's CPUID is updated is a pain, and that's what's needed for proper
FSGSBASE support.

I'll send a series, bookended by these two RFC patches, with patches to
intercept CR4 reserved bits smushed in between.  I agree there's no point in
letting the guest write LA57 directly, it's almost literally a once-per-boot
thing.  I wouldn't be surprised if intercepting it is a net win (but still
inconsequential), e.g. due to the MMU having to grab CR4 out of the VMCS.