KVM: x86: update master clock before computing kvmclock_offset

kvm master clock usually has a different frequency than the kernel boot
clock.  This is not a problem until the master clock is updated;
update uses the current kernel boot clock to compute new kvm clock,
which erases any kvm clock cycles that might have built up due to
frequency difference over a long period.

KVM_SET_CLOCK is one of places where we can safely update master clock
as the guest-visible clock is going to be shifted anyway.

The problem with current code is that it updates the kvm master clock
after updating the offset.  If the master clock was enabled before
calling KVM_SET_CLOCK, then it might have built up a significant delta
from kernel boot clock.
In the worst case, the time set by userspace would be shifted by so much
that it couldn't have been set at any point during KVM_SET_CLOCK.

To fix this, move kvm_gen_update_masterclock() before computing
kvmclock_offset, which means that the master clock and kernel boot clock
will be sufficiently close together.
Another solution would be to replace get_kvmclock_ns() with
"ktime_get_boot_ns() + ka->kvmclock_offset", which is marginally more
accurate, but would break symmetry with KVM_GET_CLOCK.

Signed-off-by: Radim Krčmář <rkrcmar@redhat.com>

arch/x86/kvm/x86.c

@@ -4188,9 +4188,15 @@ long kvm_arch_vm_ioctl(struct file *filp,
 			goto out;
 
 		r = 0;
+		/*
+		 * TODO: userspace has to take care of races with VCPU_RUN, so
+		 * kvm_gen_update_masterclock() can be cut down to locked
+		 * pvclock_update_vm_gtod_copy().
+		 */
+		kvm_gen_update_masterclock(kvm);
 		now_ns = get_kvmclock_ns(kvm);
 		kvm->arch.kvmclock_offset += user_ns.clock - now_ns;
-		kvm_gen_update_masterclock(kvm);
+		kvm_make_all_cpus_request(kvm, KVM_REQ_CLOCK_UPDATE);
 		break;
 	}
 	case KVM_GET_CLOCK: {