Hi Russell,

On Fri, Jul 25, 2014 at 05:45:19PM +0100, Russell King - ARM Linux wrote:

On Fri, Jul 25, 2014 at 06:37:46PM +0200, Maxime Ripard wrote:

quoted

Hi Vinod,

On Fri, Jul 25, 2014 at 06:42:17PM +0530, Vinod Koul wrote:

quoted

- don't use devm_request_irq(). You have irq enabled and you have killed
  tasklet. This is too racy. You need to ensure no irqs can be generated before killing
  tasklets.

Ok, would calling disable_irq before killing the tasklet an option for
you ? that would allow to keep the devm_request_irq.

That's not really an acceptable approach if you can use shared
interrupts.

We don't, but yes, I see your point.

A better alternative would be devm_free_irq() to give a definite point
that the interrupt is unregistered in the driver remove sequence.  That
allows you to keep the advantage of devm_request_irq() to clean up during
the initialisation side.

Ah, right, thanks.

An alternative approach would be to ensure that the hardware is quiesced,
and interrupts are disabled.  Then call synchronize_irq() on it, and at
that point, you should be certain that your interrupt handler should not
process any further interrupts for your device (though, in a shared
interrupt environment, it would still be called should a different device
on the shared line raise its interrupt.)

Actually, unless I'm missing something, that's pretty much what we're
doing here.

I disable all interrupts in the DMA controller, I call
synchronize_irq, and then kill the tasklet. The only interrupts I
could get are spurious, and we made sure such kind of interrupts
couldn't schedule the tasklet either.

Maxime

-- 
Maxime Ripard, Free Electrons
Embedded Linux, Kernel and Android engineering
http://free-electrons.com
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 819 bytes
Desc: Digital signature
URL: <http://lists.infradead.org/pipermail/linux-arm-kernel/attachments/20140728/5d00e9e9/attachment.sig>