xen-balloon thread using 100% of CPU, regression in 5.4.150

[Marek Marczykowski-Górecki <marmarek@xxxxxxxxxxxxxxxxxxxxxx>]

Hi,

After updating a PVH domU to 5.4.150, I see xen-balloon thread using
100% CPU (one thread).
This is a domain started with memory=maxmem=716800KiB (via libvirt). Then,
inside, I see:

# cat /sys/devices/system/xen_memory/xen_memory0/target_kb
716924
# cat /sys/devices/system/xen_memory/xen_memory0/info/current_kb
716400

Doing `cat info/current_kb > target_kb` "fixes" the issue. But still,
something is wrong - on earlier kernel (5.4.143 to be precise), it
wasn't spinning, with exactly the same values reported in sysfs. It
shouldn't run in circles if it can't get that much memory it wants. I
strongly suspect "xen/balloon: use a kernel thread instead a workqueue"
or related commit being responsible, but I haven't verified it.

This specific test is from Xen 4.8.5 (+quite a lot of patches), but I've
got report of the same issue on 4.14.3 too. Anyway, I don't think Xen
version matters here much.

I have _not_ managed to reproduce the issue on 5.10.70, nor 5.14.9. In
both cases, just after starting the domain, I see
current_kb=target_kb=716412. And writing 716924 to target_kb manually
does not cause xen-balloon thread to spin.

-- 
Best Regards,
Marek Marczykowski-Górecki
Invisible Things Lab

Attachment: signature.asc
Description: PGP signature