KVM: x86: Clamp guest MAXPHYADDR to the architectural maximum
PATCH 1/1 KVM: x86: Clamp guest MAXPHYADDR to the architectural maximum
cpuid_query_maxphyaddr() returns CPUID.0x80000008.EAX[7:0] as the guest's physical-address width without any sanity check. KVM computes the reserved GPA bits from that value, e.g. kvm_vcpu_reserved_gpa_bits_raw() is rsvd_bits(maxphyaddr, 63), and rsvd_bits() returns 0 when the bit range is empty. A VMM that populates CPUID.0x80000008.EAX above 52 therefore makes KVM stop treating the upper bits of guest PTEs as reserved: reserved-bit fault checking on guest page table walks silently degenerates, and garbage nCR3s/PDPTRs are accepted instead of faulting.
A MAXPHYADDR above 52 is architecturally impossible on x86-64; clamp the queried value to 52 so that a buggy VMM configuration degrades to a (very wide) but still meaningful physical-address width instead of disabling KVM's reserved-bit accounting.
This is hardening rather than a security fix: the actor is the VMM, which already owns the guest, so an over-wide MAXPHYADDR is not a privilege boundary crossing. The motivation is correctness, not
containment: a configuration value outside architectural limits should not silently change how KVM validates guest page tables, and clamping keeps the degenerate case well-defined instead of subtly disabling a whole class of fault checks. The unclamped read predates the introduction of the cached value; no Fixes: tag is attached for that reason.
--- arch/x86/kvm/cpuid.c | 14 ++++++++++++-- 1 file changed, 12 insertions(+), 2 deletions(-)
arch/x86/kvm/cpuid.c +12 -2
@@ -477,8 +477,18 @@ int cpuid_query_maxphyaddr(struct kvm_vcpu *vcpu)if (!best || best->eax <0x80000008 ) goto not_found; best =kvm_find_cpuid_entry (vcpu,0x80000008 ); -if (best) -return best->eax &0xff ; +if (best) { +/* +* Clamp MAXPHYADDR to the architectural maximum of 52 bits. +* A larger value is a VMM bug: KVM derives the reserved GPA +* bits from MAXPHYADDR, e.g. kvm_vcpu_reserved_gpa_bits_raw() +* is rsvd_bits(maxphyaddr, 63), and rsvd_bits() returns 0 +* when the range is empty, i.e. an over-wide MAXPHYADDR +* silently disables reserved-bit checks on guest page table +* walks. +*/ +return min_t (u32, best->eax &0xff ,52 ); + } not_found:return 36 ; }
Signature · 1 line
2.53.0
Sean Christopherson
maintainerQuoted 15 lines · Pablo Azuaje
cpuid_query_maxphyaddr()returns CPUID.0x80000008.EAX[7:0] as the guest's physical-address width without any sanity check. KVM computes the reserved GPA bits from that value, e.g.kvm_vcpu_reserved_gpa_bits_raw()isrsvd_bits(maxphyaddr, 63), andrsvd_bits()returns 0 when the bit range is empty. A VMM that populates CPUID.0x80000008.EAX above 52 therefore makes KVM stop treating the upper bits of guest PTEs as reserved: reserved-bit fault checking on guest page table walks silently degenerates, and garbage nCR3s/PDPTRs are accepted instead of faulting.A MAXPHYADDR above 52 is architecturally impossible on x86-64; clamp the queried value to 52 so that a buggy VMM configuration degrades to a (very wide) but still meaningful physical-address width instead of disabling KVM's reserved-bit accounting.
Quoted 9 lines
This is hardening rather than a security fix: the actor is the VMM, which already owns the guest, so an over-wide MAXPHYADDR is not a privilege boundary crossing. The motivation is correctness, not
containment: a configuration value outside architectural limits should not silently change how KVM validates guest page tables, and clamping keeps the degenerate case well-defined instead of subtly disabling a whole class of fault checks. The unclamped read predates the introduction of the cached value; no Fixes: tag is attached for that reason.
Unless there is a danger to the host, this isn't something I care to fix. As I said here[*] and in many other threads:
: In other words, userspace owns the vCPU model, and for good reasons. KVM needs : to allow userspace to define a sane model, but with *very* few exceptions, KVM : should not try to "help" userspace by stuffing bits.
:
: Pretty much everytime KVM tries to help, it causes problems. E.g. initializing : perf_capabilities to kvm_caps.supported_perf_cap seems like a good thing, except : it presents a bogus model if userspace decides to not enumerate a vPMU to the : guest (Aaron was allegedly going to send a patch for this...).
There are myriad ways for the VMM to misconfigure the guest, and there are also legitimate reasons why userspace may want to do something that KVM thinks is odd.
If for some reason we do need to "fix" this, I would want to reject KVM_SET_CPUID{,2}, not silently clobber the vCPU model. But that's a big "if", because AFAICT there is no danger to the host.