Re: INFO: possible circular locking dependency at cleanup_workqueue_thread

[Oleg Nesterov <oleg@redhat.com>]

On 05/19, Johannes Berg wrote:
>
> On Tue, 2009-05-19 at 14:00 +0200, Oleg Nesterov wrote:
>
> > > I'm not familiar enough with the code -- but what are we really trying
> > > to do in CPU_POST_DEAD? It seems to me that at that time things must
> > > already be off the CPU, so ...?
> >
> > Yes, this cpu is dead, we should do cleanup_workqueue_thread() to kill
> > cwq->thread.
> >
> > > On the other hand that calls
> > > flush_cpu_workqueue() so it seems it would actually wait for the work to
> > > be executed on some other CPU, within the CPU_POST_DEAD notification?
> >
> > Yes. Because we can't just kill cwq->thread, we can have the pending
> > work_structs so we have to flush.
> >
> > Why can't we move these works to another CPU? We can, but this doesn't
> > really help. Because in any case we should at least wait for
> > cwq->current_work to complete.
> >
> > Why do we use CPU_POST_DEAD, and not (say) CPU_DEAD to flush/kill ?
> > Because work->func() can sleep in get_online_cpus(), we can't flush
> > until we drop cpu_hotplug.lock.
>
> Right. But exactly this happens in the hibernate case --

not sure I understand your "exactly this" ;)

But your explanation of the deadlock below looks great!

> the hibernate
> code calls kernel/cpu.c:disable_nonboot_cpus() which calls _cpu_down()
> which calls raw_notifier_call_chain(&cpu_chain, CPU_POST_DEAD... Sadly,
> it does so while holding the cpu_add_remove_lock, which is happens to
> have the dependencies outlined in the original email...
>
> The same happens in cpu_down() (without leading _) which you can trigger
> from sysfs by manually removing the CPU, so it's not hibernate specific.

except I don't understand how cpu_add_remove_lock makes the difference...
And thus I can't understand why cpu_down() (called lockless) have the
same problems. Please see below.

> Anyway, you can have a deadlock like this:
>
> CPU 3			CPU 2				CPU 1
> 							suspend/hibernate
> 			something:
> 			rtnl_lock()			device_pm_lock()
> 							-> mutex_lock(&dpm_list_mtx)
>
> 			mutex_lock(&dpm_list_mtx)
>
> linkwatch_work
>  -> rtnl_lock()
> 							disable_nonboot_cpus()

let's suppose disable_nonboot_cpus() does not take cpu_add_remove_lock,

> 							-> flush CPU 3 workqueue

in this case the deadlock is still here?

We can't flush because we hold the lock (dpm_list_mtx) which depends
on another lock taken by work->func(), the "classical" bug with flush.

No?

Oleg.