Re: [REGRESSION 7.2, BISECTED] KVM: x86: Starting Windows guest triggers UBSAN: array-index-out-of-bounds

[Sean Christopherson <seanjc@google.com>]

On Wed, Oct 07, 2026, Sean Christopherson wrote:
> On Wed, Oct 07, 2026, Paolo Bonzini wrote:
> > On Tue, Oct 6, 2026 at 7:37 PM SimonP <simonp.git@mailbox.org> wrote:
> > > Hello,
> > >
> > > See subject. Happens in permission_fault() on line:
> > >
> > >         fault = (mmu->permissions[index] >> pte_access) & 1;
> > >
> > > Some output from a printk I added, not spammed but keeps happening:
> > >
> > > kvm: permission_fault(): pfec = 70, index = 35, not_smap = 0
> > 
> > Can you confirm you are using nested virtualization, and if so what is
> > the L1 hypervisor and L2 guest combination? Also what is the processor
> > model?
> > 
> > The issue is weird; it is caused by the SS bit which is apparently set
> > in the #NPF exitcode; but it shouldn't be unless bit 4 is set in the
> > misc_ctl field, and KVM never sets it. Can you check if the attached
> > patch fixes it (modulo the fact that it shouldn't happen in the first
> > place)?
> 
> ...
> 
> > > kvm: permission_fault(): pfec = 70, index = 35, not_smap = 0
> > > kvm: permission_fault(): pfec = 70, index = 35, not_smap = 0
> > > kvm: permission_fault(): pfec = 70, index = 35, not_smap = 0
> > > kvm: permission_fault(): pfec = 68, index = 34, not_smap = 0
> > > kvm: permission_fault(): pfec = 68, index = 34, not_smap = 0
> 
> It's not just SS, it's also PK (which shouldn't happen either) and RSVD, which
> *really* shouldn't happen.  In the page fault path, KVM explicitly clears RSVD
> from error code that becomes @access and then pfec, so I'm completely baffled.
> Nothing in walk_addr_generic() writes to @access, so the only thing I can think
> of is that we're clobbering the stack?
> 
> 	r = FNAME(walk_addr)(&walker, vcpu, fault->addr,
> 			     fault->error_code & ~PFERR_RSVD_MASK);

Paolo pointed out offlist that the printks are in decimal, not hex.  *sigh*

So yeah, 70 = 0x46 = SS | USER | WRITE.  Which is way more sane than SS | PK | RSVD. :-D

While I'm here, the theory is that hardware is setting the SS bit because the
faulting access really was a userspace shadow stack access.  Which was obviously
unexpected by us, but might be working as intended?