On Fri, Jan 14, 2022 at 03:09:55PM +0100, Peter Zijlstra wrote:

quoted

I think the assumption is based on the idea that a process
using UMCG will get affined to N CPUs, will have N servers and
a number of workers, and they will all happily cooperate and not
get any extra threads running.

Of course the pretty picture was not completely true, as the unblocked
tasks do consume extra threads in the kernel, though never in the
userspace.

Right, there is some unmanaged time anyway.

Also, since we force wake to the same CPU, and overlapping runtime is
'short', they should all stick to the same CPU, even if we don't pin.