Re: [PATCH] LIS3LV02Dx Accelerometer driver (take 4)

[Andrew Morton <akpm@linux-foundation.org>]

On Wed, 22 Oct 2008 00:13:42 +0200
Eric Piel <eric.piel@tremplin-utc.net> wrote:

> Andrew Morton schreef:
> >> +static void lis3lv02d_poweroff_timeout(unsigned long data)
> >> +{
> >> +	struct acpi_lis3lv02d *dev = (void *)data;
> >> +
> >> +	up(&dev->poff_sem);
> >> +	lis3lv02d_poweroff(dev->device->handle);
> >> +	down(&dev->poff_sem);
> > 
> > eek, no, we cannot down a semaphore from a timer handler!  It will lead
> > to might_sleep() warnings, scheduling-in-atomic warnings and kernel
> > deadlocks.
> Ooh... What happened is that I wrote the other functions first. They
> indirectly use acpi_evaluate_integer(), which can sleep. So I avoided
> spin_locks, and decided to use a semaphore. But now there is a semaphore
> in a timer (which is not allowed to sleep).
> 
> It seems the semaphore usage is compulsory due to acpi, so I'm planning
> to change the timer into a scheduled work on a workqueue (with
> queue_delayed_work(),  and cancel_delayed_work_sync()). This workqueue
> could be the same as the one for the joystick input. Does that look the
> right way to do it?

yep, that's generally a safe transformation.

You might also be able to use keventd, via scheule_delayed_work().