Re: INFO: possible circular locking dependency atcleanup_workqueue_thread

[Johannes Berg]

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 -- 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.

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()
							-> flush CPU 3 workqueue

johannes

Attachment: signature.asc
Description: This is a digitally signed message part