Re: [PATCH] pcf50633: use a dedicated workqueue for irq processing

[Anton Vorontsov <avorontsov@ru.mvista.com>]

On Fri, Jul 31, 2009 at 01:23:13PM +0400, Paul Fertser wrote:
> On Thu, Jul 30, 2009 at 03:34:25PM -0700, Andrew Morton wrote:
> > On Tue, 28 Jul 2009 00:41:15 +0400
> > Paul Fertser <fercerpav@gmail.com> wrote:
> > 
> > > Using the default kernel "events" workqueue causes problems with
> > > synchronous adc readings if initiated from some task on the same
> > > workqueue.
> > > 
> > > I had a deadlock trying to use pcf50633_adc_sync_read from a
> > > power_supply class driver because the reading was initiated from the
> > > workqueue and it waited for the irq processing to complete (to get the
> > > result) and that was put on the same workqueue.
> > 
> > I don't get it.
> > 
> > Do you meant that pcf50633_adc_sync_read() was called via a
> > schedule_work() handler?  If so, wasn't that a bug?
> 
> Andrew, i think it is called this way:
> 
> power_supply_changed -> schedule_work(&psy->changed_work) ->
> power_supply_changed_work -> kobject_uevent(&psy->dev->kobj, KOBJ_CHANGE)
> -> power_supply_uevent -> power_supply_show_property -> get_property ->
> bat->pdata->get_voltage -> pcf50633_adc_sync_read
> 
> I assumed that since power_supply class is well-established and widely used
> that's an intended behaviour. CCing relevant maintainers to get their
> opinions.

The thing is that power_supply_changed() is supposed to be callable
from any context (including hard IRQs), so we use schedule_work().

I see the deadlock problem, and it could be fixed by a dedicated
workqueue either in the pcf50633 driver, or power supply class.
I don't mind either way.

Another option is to implement power_supply_changed_can_sleep(),
which won't use schedule_work(), and so we won't need another
thread.

Thanks,

-- 
Anton Vorontsov
email: cbouatmailru@gmail.com
irc://irc.freenode.net/bd2