Re: [PATCH v3] cgroup: avoid flushing global workqueue in cgroup1_pidlist_destroy_all

[Junnan Zhang]

Hi Tejun, Michal,

> I'd rather not touch them at all and eventually retire the whole v1
> interface.

Understood, and I have no objection to Michal's approach as the
long-term direction -- routing v1 reads through css_task_iter removes
the pidlist machinery altogether, including the destroy workqueue
behind this hang.

May I ask for some guidance on the near term, though? cgroup v1 is
still the default on widely deployed enterprise distros and will
likely remain in production use for years to come, and this hung
task was observed on a production Kubernetes node. This patch (v3)
is a pure maintenance fix: it changes no user-visible behavior, adds
no interface, and is stable-backportable. It would also remain
relevant for the cgroup_v1_sorted=true path if Michal's change
lands.

Would you consider taking it as a minimal fix? Or, if the consensus
is to leave v1 completely untouched, what would be the recommended
course for production systems that keep hitting this?

Thanks,
Junnan