KVM: nVMX: Document that ignoring memory failures for VMCLEAR is deliberate

Explicitly drop the result of kvm_vcpu_write_guest() when writing the
"launch state" as part of VMCLEAR emulation, and add a comment to call
out that KVM's behavior is architecturally valid.  Intel's pseudocode
effectively says that VMCLEAR is a nop if the target VMCS address isn't
in memory, e.g. if the address points at MMIO.

Add a FIXME to call out that suppressing failures on __copy_to_user() is
wrong, as memory (a memslot) does exist in that case.  Punt the issue to
the future as open coding kvm_vcpu_write_guest() just to make sure the
guest dies with -EFAULT isn't worth the extra complexity.  The flaw will
need to be addressed if KVM ever does something intelligent on uaccess
failures, e.g. to support post-copy demand paging, but in that case KVM
will need a more thorough overhaul, i.e. VMCLEAR shouldn't need to open
code a core KVM helper.

No functional change intended.

Reported-by: coverity-bot <keescook+coverity-bot@chromium.org>
Addresses-Coverity-ID: 1527765 ("Error handling issues")
Fixes: 587d7e72aedc ("kvm: nVMX: VMCLEAR should not cause the vCPU to shut down")
Cc: Jim Mattson <jmattson@google.com>
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20221220154224.526568-1-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

diff --git a/arch/x86/kvm/vmx/nested.c b/arch/x86/kvm/vmx/nested.c
index b6f4411..f18f3a9 100644 (file)
--- a/arch/x86/kvm/vmx/nested.c
+++ b/arch/x86/kvm/vmx/nested.c
@@ -5296,10 +5296,19 @@ static int handle_vmclear(struct kvm_vcpu *vcpu)
                if (vmptr == vmx->nested.current_vmptr)
                        nested_release_vmcs12(vcpu);
 
-               kvm_vcpu_write_guest(vcpu,
-                                    vmptr + offsetof(struct vmcs12,
-                                                     launch_state),
-                                    &zero, sizeof(zero));
+               /*
+                * Silently ignore memory errors on VMCLEAR, Intel's pseudocode
+                * for VMCLEAR includes a "ensure that data for VMCS referenced
+                * by the operand is in memory" clause that guards writes to
+                * memory, i.e. doing nothing for I/O is architecturally valid.
+                *
+                * FIXME: Suppress failures if and only if no memslot is found,
+                * i.e. exit to userspace if __copy_to_user() fails.
+                */
+               (void)kvm_vcpu_write_guest(vcpu,
+                                          vmptr + offsetof(struct vmcs12,
+                                                           launch_state),
+                                          &zero, sizeof(zero));
        } else if (vmx->nested.hv_evmcs && vmptr == vmx->nested.hv_evmcs_vmptr) {
                nested_release_evmcs(vcpu);
        }