On Thu, Aug 15, 2024 at 12:35 AM Sean Christopherson <seanjc@xxxxxxxxxx> wrote: > > On Wed, Aug 14, 2024, Suleiman Souhlal wrote: > > > With the proposed approach, the steal time page would get copied to the new > > host and everything would keep working correctly, with the exception of a > > possible host suspend happening between when the migration started and when > > it finishes, not being reflected post-migration. That seems like a > > reasonable compromise. > > Maybe, but I'm not keen on sweeping this under the rug. Ignoring issues because > they'll "never" happen has bitten KVM more than once. > > At the absolute bare minimum, the flaw needs to be documented, with a suggested > workaround provided (do KVM on all vCPUs before migrating after suspend), e.g. > so that userspace can workaround the issue in the unlikely scenario userspace > does suspend+resume, saves/restores a VM, *and* cares about steal-time. I can write a comment in record_steal_time() that describes the scenario, mention it in the commit message and add something to the steal time part of Documentation/virt/kvm/x86/msr.rst. -- Suleiman
