Re: [RFC][PATCH] PM: Introduce new top level suspend and hibernation callbacks (rev. 3)

[Rafael J. Wysocki]

On Wednesday, 26 of March 2008, Alan Stern wrote:
> On Wed, 26 Mar 2008, Rafael J. Wysocki wrote:
> 
> > On Wednesday, 26 of March 2008, Alan Stern wrote:
> > > On Tue, 25 Mar 2008, Rafael J. Wysocki wrote:
> > > 
> > > > > I just thought of another problem.  At the point where
> > > > > local_irq_disable() is called, in between device_suspend() and
> > > > > device_power_down(), it is possible in a preemptible kernel that
> > > > > another task is holding dpm_list_mtx and is in the middle of updating
> > > > > the list pointers.  This would mess up the traversal in
> > > > > device_power_down().
> > > > > 
> > > > > I'm not sure about the best way to prevent this.  Is it legal to call
> > > > > unlock_mutex() while interrupts or preemption are disabled?
> > > > 
> > > > Well, I think it is, but I'm not sure how that can help.
> > > > 
> > > > To prevent the race from happening, we can lock dpm_list_mtx before switching
> > > > interrupts off in kernel/power/main.c:suspend_enter() and analogously in
> > > > kernel/power/disk.c .
> > > 
> > > That's right.  And once interrupts are turned off you should unlock 
> > > dpm_list_mtx again, in case a noirq method wants to unregister a 
> > > device.
> > 
> > Why would a noirq method want to do that?  IMO, it's not a big deal if noirq
> > methods are not allowed to unregister devices.
> 
> Okay, that's fine.  It keeps things simple.
> 
> > > Hence my question: Is it legal to call unlock_mutex() while interrupts are
> > > disabled? 
> > 
> > Well, I suspect that will confuse lockdep quite a bit.  Otherwise, I don't see
> > a problem with it (it's just changing the value of a shared variable after
> > all).
> 
> Then you have your answer.  Perhaps have device_suspend() exit with the
> mutex held and have device_resume() release it (with appropriate
> handling for error situations, of course).

That wouldn't work, because enable_nonboot_cpus() is called before
device_resume() and the notifiers in there may want to unregister devices
if some CPUs fail to go online.

I added two accessor functions device_pm_lock() and device_pm_unlock()
to be called just prior to disabling interrupts and right after enabling them,
respectively, in the higher-level PM core (ie. kernel/power/main(disk).c).

Thanks,
Rafael
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/