Re: BUG: using __this_cpu_read() in preemptible  code: mm_percpu_wq/7
From: Heiko Carstens
Date: Wed Jul 12 2017 - 12:13:52 EST
On Wed, Jul 12, 2017 at 10:44:54AM -0500, Christopher Lameter wrote:
> On Wed, 7 Jun 2017, Andre Wild wrote:
> > I'm currently seeing the following message running kernel version 4.11.0.
> > It looks like it was introduced with the patch
> > 4037d452202e34214e8a939fa5621b2b3bbb45b7.
> A 2007 patch? At that point we did not have __this_cpu_read() nor
> refresh_cpu_vmstats.... Is this on s390 or some such architecture?
It is on s390, but after I looked into the code a bit deeper the mentioned
patch doesn't seem to be the problem.
My initial thought was a missing preempt_disable() / preempt_enable() pair,
but that can't be the problem, since the code is executed on a per-cpu
> > Can you please take a look at this problem?
> Could you give me a bit more context?
> > [Tue Jun 6 15:27:03 2017] BUG: using __this_cpu_read() in preemptible
> >  code: mm_percpu_wq/7
> > [Tue Jun 6 15:27:03 2017] caller is refresh_cpu_vm_stats+0x198/0x3d8
> > [Tue Jun 6 15:27:03 2017] CPU: 0 PID: 7 Comm: mm_percpu_wq Tainted: G
> > W 4.11.0-20170529.0.ae409ab.224a322.fc25.s390xdefault #1
> > [Tue Jun 6 15:27:03 2017] Workqueue: mm_percpu_wq vmstat_update
> It is run in preemptible mode but this from a kworker
> context so the processor cannot change (see vmstat_refresh()).
> Even on s390 or so this should be fine.
Yes, it *should* be fine. The only unusual thing here is that this happens
during quite a lot of cpu hotplug operations. So even though the workqueue
code should be able to handle cpu hotplug correctly, my best guess is that
current->cpus_allowed is not cpumask_of(this_cpu) for some reason.
That would be this check within lib/smp_processor_id.c:check_preemption_disabled()
if (cpumask_equal(¤t->cpus_allowed, cpumask_of(this_cpu)))
I changed the code to simply panic, so I can look into a dump to figure out
what actually does cause the warning. As soon as Andre finds some time to
reproduce this we will come back to you.