On 4/5/2024 4:01 PM, Sean Christopherson wrote: > On Fri, Mar 22, 2024, Zide Chen wrote: >> Currently, the migration worker delays 1-10 us, assuming that one >> KVM_RUN iteration only takes a few microseconds. But if C-state exit >> latencies are large enough, for example, hundreds or even thousands >> of microseconds on server CPUs, it may happen that it's not able to >> bring the target CPU out of C-state before the migration worker starts >> to migrate it to the next CPU. >> >> If the system workload is light, most CPUs could be at a certain level >> of C-state, and the vCPU thread may waste milliseconds before it can >> actually migrate to a new CPU. > > Well fudge. That's definitely not on my bingo sheet. > >> Thus, the tests may be inefficient in such systems, and in some cases >> it may fail the migration/KVM_RUN ratio sanity check. >> >> Since we are not able to turn off the cpuidle sub-system in run time, >> this patch creates an idle thread on every CPU to prevent them from >> entering C-states. > > First off, huge thanks for debugging this! That must have been quite the task > (no pun intended). > > While spinning up threads on every CPU is a clever way to ensure they don't go > into a deep sleep state, I'm not exactly excited about the idea of putting every > reachable CPU into a busy loop. And while this doesn't add _that_ much complexity, > I'm not sure the benefit (preserving the assert for all systems) is worth it. I > also don't want to arbitrarily prevent idle task (as in, the kernel's idle task) > interactions. E.g. it's highly (highly) unlikely, but not impossible for there > to be a bug that's unique to idle tasks, or C-states, or other edge case. > > Are there any metrics/stats that can be (easily) checked to grant an exception > to the sanity check? That's a very hand-wavy question, as I'm not even sure what > type of stat we'd want to look at. Actual runtime of a task, maybe? > > If that's not easy, what if we add an off-by-default command line option to skip > the sanity check? I was resistant to simply deleting the assert in the past, but > that was mainly because I didn't want to delete it without understanding what was > causing problems. That would allow CI environments to opt-out as needed, while > still keeping the sanity check alive for enough systems to make it useful. Sorry for not replying earlier. I overlooked your email from my inbox. :) Alternative to the busy loop, how about using the /dev/cpu_dma_latency interface to disable c-states (I wish I had learned this before writing the initial patch)? The good thing is it can do automatic cleanup when it closes the fd. The reason why I still think of disabling c-states is, even in the low c-states exit latency setup, it can still increase the chances of successful migration. Otherwise, I can implement a command line option to skip the sanity check, as I was not able to find out a metrics/stats that is easy and reliable to indicate that the scheduler is not able to wake up the target CPU before the task is scheduled to another CPU.
