weird /proc/stat output with newer (4.1, 4.2) kernels in kvm guest on 3.10 host?

Hi,

I'm running a 3.10-based host with libvirt 1.2.12 and qemu 2.2.

Running a Fedora23 cloud image as a guest, the "cpu" lines in /proc/stat seem to be hardly changing:

[fedora@fedora23 boot]$ uptime
17:01:50 up 44 min, 1 user, load average: 3.00, 2.99, 2.79
[fedora@fedora23 boot]$ grep cpu /proc/stat
cpu 13 0 124 65880 165 0 0 23088 0 0
cpu0 1 0 4 21967 67 0 0 21866 0 0
cpu1 8 0 52 21877 66 0 0 623 0 0
cpu2 4 0 67 22034 30 0 0 598 0 0

Running an Ubuntu 15.10 cloud image as a guest, everything looks fine:

ubuntu@ubuntu1510:/boot$ uptime
 17:02:28 up 28 min,  1 user,  load average: 3.00, 2.98, 2.42
ubuntu@ubuntu1510:/boot$ grep cpu /proc/stat
cpu  3932 130 443061 46116 257 0 5 12974 0 0
cpu0 1296 130 139886 14960 195 0 1 12343 0 0
cpu1 1330 0 151580 15570 35 0 0 315 0 0
cpu2 1305 0 151595 15586 26 0 4 316 0 0

In both cases the current_clocksource is kvm-clock. Not sure if it's relevent, but Fedora has "CONFIG_NO_HZ_FULL=y" and "CONFIG_NO_HZ_FULL_ALL is not set", and does not specify "nohz_full=" in the boot parameters. Ubuntu has "CONFIG_NO_HZ_FULL is not set".

Anyone have any ideas what might be going on?

Chris

Quick update (sorry it's not threaded, I'm not subscribed to the list).

The problem seems to be related to VIRT_CPU_ACCOUNTING being enabled in the guest. If I switch to TICK_CPU_ACCOUNTING or IRQ_TIME_ACCOUNTING then the problem goes away.

I'm digging into the code now. Anyone have any tips on debugging why VIRT_CPU_ACCOUNTING is misbehaving in the guest?

Thanks,
Chris