Re: [PATCH] cpusets: rework guarantee_online_cpus() to fix deadlock with cpu_down()

[Oleg Nesterov <oleg@redhat.com>]

On 08/04, Lai Jiangshan wrote:
>
> Oleg Nesterov wrote:
> > On 08/02, Lai Jiangshan wrote:
> >> Oleg Nesterov wrote:
> >>
> >>> 	- do NOT scan cs->parent cpusets. If there are no online CPUs in
> >>> 	  cs->cpus_allowed, we use cpu_online_mask. This is only possible
> >>> 	  when we are called by cpu_down() hooks, in that case
> >>> 	  cpuset_track_online_cpus(CPU_DEAD) will fix things later.
> >>>
> >>
> >> We must scan cs->parent cpusets.
> >> A task is constrained by a cpuset,
> >
> > Yes, the task esacpes its cpuset. With or without this patch.
> > Because cs->cpus_allowed has no online CPUs.
> >
> >> it must be constrained
> >> this cpuset's parent too.
> >
> > It will be constained again, after scan_for_empty_cpusets(), no?
>
> cpuset_cpus_allowed() is not only  used for CPU offline.
>
> sched_setaffinity() also uses it.

Sure. And it must take get_online_cpus() to avoid the races with hotplug.

> The task will not be constained again.

It will be constrained, and we don't need to scan cs->parent.

Because, since we are protected from cpu_down(), cs->cpus_allowed must
have online CPUs. update_cpumask() doesn't allow the empty cpumasks.

> Or I missed something.

Or me ;)

Oleg.