KVM: x86: fix L1TF's MMIO GFN calculation

KVM: x86: fix L1TF's MMIO GFN calculation

One defense against L1TF in KVM is to always set the upper five bits
of the *legal* physical address in the SPTEs for non-present and
reserved SPTEs, e.g. MMIO SPTEs.  In the MMIO case, the GFN of the
MMIO SPTE may overlap with the upper five bits that are being usurped
to defend against L1TF.  To preserve the GFN, the bits of the GFN that
overlap with the repurposed bits are shifted left into the reserved
bits, i.e. the GFN in the SPTE will be split into high and low parts.
When retrieving the GFN from the MMIO SPTE, e.g. to check for an MMIO
access, get_mmio_spte_gfn() unshifts the affected bits and restores
the original GFN for comparison.  Unfortunately, get_mmio_spte_gfn()
neglects to mask off the reserved bits in the SPTE that were used to
store the upper chunk of the GFN.  As a result, KVM fails to detect
MMIO accesses whose GPA overlaps the repurprosed bits, which in turn
causes guest panics and hangs.

Fix the bug by generating a mask that covers the lower chunk of the
GFN, i.e. the bits that aren't shifted by the L1TF mitigation.  The
alternative approach would be to explicitly zero the five reserved
bits that are used to store the upper chunk of the GFN, but that
requires additional run-time computation and makes an already-ugly
bit of code even more inscrutable.

I considered adding a WARN_ON_ONCE(low_phys_bits-1 <= PAGE_SHIFT) to
warn if GENMASK_ULL() generated a nonsensical value, but that seemed
silly since that would mean a system that supports VMX has less than
18 bits of physical address space...

Reported-by: Sakari Ailus <sakari.ailus@iki.fi>
Fixes: d9b47449c1a1 ("kvm: x86: Set highest physical address bits in non-present/reserved SPTEs")
Cc: Junaid Shahid <junaids@google.com>
Cc: Jim Mattson <jmattson@google.com>
Cc: stable@vger.kernel.org
Reviewed-by: Junaid Shahid <junaids@google.com>
Tested-by: Sakari Ailus <sakari.ailus@linux.intel.com>
Signed-off-by: Sean Christopherson <sean.j.christopherson@intel.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>