fix(kvmclock): don't pass KVM_CLOCK_REALTIME to KVM_SET_CLOCK

When we pass KVM_CLOCK_REALTIME to KVM_SET_CLOCK, the kernel steps the
monotonic clock by the elapsed wallclock time. This is ok for live
migration, but in the case of snapshots that were take long ago it might
cause issues to guest userspace.

Change the snapshot restore logic to not clear the KVM_CLOCK_REALTIME
bit when restoring kvmclock. Wallclock time will then be handled by
userspace (chronyd or otherwise).

Signed-off-by: Babis Chalios <babis.chalios@e2b.dev>

Author: bchalios

src/vmm/src/arch/x86_64/vm.rs

@@ -4,8 +4,7 @@
 use std::fmt;
 
 use kvm_bindings::{
-    KVM_CLOCK_TSC_STABLE, KVM_IRQCHIP_IOAPIC, KVM_IRQCHIP_PIC_MASTER, KVM_IRQCHIP_PIC_SLAVE,
-    KVM_PIT_SPEAKER_DUMMY, MsrList, kvm_clock_data, kvm_irqchip, kvm_pit_config, kvm_pit_state2,
+    KVM_CLOCK_REALTIME, KVM_CLOCK_TSC_STABLE, KVM_IRQCHIP_IOAPIC, KVM_IRQCHIP_PIC_MASTER, KVM_IRQCHIP_PIC_SLAVE, KVM_PIT_SPEAKER_DUMMY, MsrList, kvm_clock_data, kvm_irqchip, kvm_pit_config, kvm_pit_state2
 };
 use kvm_ioctls::Cap;
 use serde::{Deserialize, Serialize};
@@ -122,8 +121,13 @@ impl ArchVm {
         self.fd()
             .set_pit2(&state.pitstate)
             .map_err(ArchVmError::SetPit2)?;
+
+        // Do not pass KVM_CLOCK_REALTIME to the ioctl(). This causes the kernel to
+        // adjust the MONOTONIC clock which might cause issues for the guest userspace.
+        let mut clock = state.clock;
+        clock.flags &= !KVM_CLOCK_REALTIME;
         self.fd()
-            .set_clock(&state.clock)
+            .set_clock(&clock)
             .map_err(ArchVmError::SetClock)?;
         self.fd()
             .set_irqchip(&state.pic_master)