On Fri, Aug 27, 2021, Mathieu Desnoyers wrote:
So there are effectively three reasons we want a delay:
To allow sched_setaffinity() to coincide with ioctl(KVM_RUN) before KVM can enter the guest so that the guest doesn't need an arch-specific VM-Exit source.
To let ioctl(KVM_RUN) make its way back to the test before the next round of migration.
To ensure the read-side can make forward progress, e.g. if sched_getcpu() involves a syscall.
After looking at KVM for arm64 and s390, #1 is a bit tenuous because x86 is the only arch that currently uses xfer_to_guest_mode_work(), i.e. the test could be tweaked to be overtly x86-specific. But since a delay is needed for #2 and #3, I'd prefer to rely on it for #1 as well in the hopes that this test provides coverage for arm64 and/or s390 if they're ever converted to use the common xfer_to_guest_mode_work().
Now that we have this understanding of why we need the delay, it would be good to write this down in a comment within the test.
Ya, I'll get a new version out next week.
Does it reproduce if we randomize the delay to have it picked randomly from 0us to 100us (with 1us step) ? It would remove a lot of the needs for arch-specific magic delay value.
My less-than-scientific testing shows that it can reproduce at delays up to ~500us, but above ~10us the reproducibility starts to drop. The bug still reproduces reliably, it just takes more iterations, and obviously the test runs a bit slower.
Any objection to using a 1-10us delay, e.g. a simple usleep((i % 10) + 1)?