On 21/05/21 01:03, Jim Mattson wrote:
When the VMCS12 posted interrupt descriptor isn't backed by an L1
memslot, kvm will launch vmcs02 with a stale posted interrupt
descriptor. Before commit 6beb7bd52e48 ("kvm: nVMX: Refactor
nested_get_vmcs12_pages()"), kvm would have silently disabled the
VMCS02 "process posted interrupts" VM-execution control. Both
behaviors are wrong, though the use-after-free is more egregious.
Empirical tests on actual hardware reveal that a posted interrupt
descriptor without any backing memory/device has PCI bus error
semantics (reads return all 1's and writes are discarded). However,
kvm can't tell an unbacked address from an MMIO address. Normally, kvm
would ask userspace for an MMIO completion, but that's complicated for
a posted interrupt descriptor access. There are already a number of
cases where kvm bails out to userspace with KVM_INTERNAL_ERROR via
kvm_handle_memory_failure, so that seems like the best route to take.
It would be relatively easy to invoke kvm_handle_memory_failure at
emulated VM-entry, but that approach would break existing
kvm-unit-tests. Moreover, the issue doesn't really come up until the
vCPU--in virtualized VMX non-root operation--received the posted
interrupt notification vector indicated in its VMCS12.
Sadly, it's really hard to arrange for an exit to userspace from
vmx_complete_nested_posted_interrupt, which is where kvm actually
needs to access the unbacked PID. Initially, I added a new kvm request
for a userspace exit on the next guest entry, but Sean hated that
approach. Based on his suggestion, I added the plumbing to get back
out to userspace in the event of an error in
vmx_complete_nested_posted_interrupt. This works in the case of an
unbacked PID, but it doesn't work quite as well in the case of an
unbacked virtual APIC page (another case where kvm was happy to just
silently ignore the problem and attempt to muddle its way through.) In
that case, this series is an incremental improvement, but it's not a
complete fix.
Jim Mattson (12):
KVM: x86: Remove guest mode check from kvm_check_nested_events
KVM: x86: Wake up a vCPU when kvm_check_nested_events fails
KVM: nVMX: Add a return code to vmx_complete_nested_posted_interrupt
KVM: x86: Add a return code to inject_pending_event
KVM: x86: Add a return code to kvm_apic_accept_events
KVM: nVMX: Fail on MMIO completion for nested posted interrupts
KVM: nVMX: Disable vmcs02 posted interrupts if vmcs12 PID isn't
mappable
KVM: selftests: Move APIC definitions into a separate file
KVM: selftests: Hoist APIC functions out of individual tests
KVM: selftests: Introduce x2APIC register manipulation functions
KVM: selftests: Introduce prepare_tpr_shadow
KVM: selftests: Add a test of an unbacked nested PI descriptor
Patch 2 is the only one that seems wrong to me, and I am actually not
sure why it is part of this series. It seems to me that it overlaps
"kvm: x86: move srcu lock out of kvm_vcpu_check_block", for which both
Sean and I had some discussions on how to remove the side effects that
kvm_check_nested_events has on kvm_vcpu_is_runn{able,ing}.
Otherwise looks good, and I've already queued patches 8-10. Holler if I
should just apply patches 1 and 3-12.
Paolo
