Re: [PATCH 6/7] virt: steal_governor: restart sampling immediately after CPU hotplug

From: Shrikanth Hegde

Date: Fri Nov 20 2026 - 03:51:14 EST




On 11/19/26 11:09 PM, Yury Norov wrote:
On Thu, Nov 19, 2026 at 11:36:46AM +0530, Shrikanth Hegde wrote:
Hi Yury.

On 11/14/26 12:12 AM, Yury Norov wrote:
CPU hotplug between samples can distort the calculated steal ratio because
the governor divides accumulated steal time by the current active CPU
count, even when that count changed during the sampling interval.

For example, assume a constant 3% steal rate on each CPU, with four active
CPUs for 900ms followed by two active CPUs for 100ms. The accumulated
steal time is:

4 * 900ms * 3% + 2 * 100ms * 3% = 108ms + 6ms = 114ms


This example is not realistic. If you had 3% steal with 4 CPUs, and if they become 2,
your steal time will definitely increase. It will likely double. So
governor likely calculates 6%.

You say it's not realistic, and then you confirm the math. So
realistic or not?

I do not expect the steal time to remain consistent across after hotplug
completes.

Instead of speculating, we should to get the numbers from the actual systems.

Let me see if i get some time to get it w.r.t to powerVM implementation.


On my VM, there are 4 or 8 vCPUs, and the steal time hovers around 3%
normally. Offlining half of them triggers steal governor pretty
consistently.
But yes, you can have realistic examples, where governor may calculate
the values which goes slightly above or below. But IMHO those examples are very
limited. Even in the above examples, the steal governor was supposed to trigger
high threshold since the steal time with 2 CPUs is 6%. So if not now, in the next
delta it would have triggered it anyways.

Sorry, I don't understand this. The governor is not supposed to
trigger decrease_preferred_cpus() on offline event. If some system
event skews steal ratio, the governor must adjust to that, or simply
skip the iteration. This patch implements the 2nd approach, which is
simpler IMO.

Some examples are really corner cases. Maybe like

60 CPUs 3% steal, and then 40 CPUs 5.1% steal = 4.55%
(Below threshold, but should have been high). One interval missed. But if you see
large duration was in low threshold, so it is agreeable to miss it.

Don't understand this either. Can you elaborate on your case and math
please?
The elapsed time is 1000ms. Using the final count of two active CPUs, the
governor calculates:

114ms / (2 * 1000ms) = 5.7%

This exceeds the default 5% high threshold, even though the actual steal
rate stayed at 3%. Using the CPU counts over the whole interval gives:

114ms / (4 * 900ms + 2 * 100ms) = 3%

Checking num_online_cpus() doesn't account for cases where CPUs go online
and offline such that the total number of online CPUs doesn't change. So,
implement the sampling restart on hotplug event.


I am not sure if we really need to stress about hotplug cases. IIUC, when a user
does hotplug, systems are usually idle. (Atleast that is what we have observed in
real life use cases). And hotplug is really user driven, so user knows system is going
through some transitions and will be stable soon.

The follow-up series brings a real counter example: Android daemon
running on Qualcomm SoC offlines CPUs on thermal event. So, it's
automatic and the offlined CPU is likely not idling.


I wasn't aware of that. So is it userspace daemon or kernel module/driver?
Thanks,
Yury