On Wed, Jul 31, 2013 at 10:15:12AM -0400, Tejun Heo wrote:

hello,

On Wed, Jul 31, 2013 at 09:55:26PM +0800, Peter Chen wrote:

quoted

I think the main point is we should allocate managed resource which is used
at interrupt handler before devm_request_irq, and all resources used
at interrupt
handler should be managed.

If we use non-managed resource at interrupt handler, but using managed interrupt
handler, things still will go wrong if there is an odd (unexpected)
interrupt after
we finish deactivation at removal.

In general, applying devm partially isn't a good idea.  It's very easy
to get into trouble thanks to release order dependency which isn't
immediately noticeable and there have been actual bugs caused by that.
The strategies which seem to work are either

* Convert everything over to devm by wrapping deactivation in a devres
  callback too.  As long as your init sequence is sane (ie. irq
  shouldn't be request before things used by irq are ready).

* Allocate all resources using devres but shut down the execution
  engine in the remove_one().  Again, as all releases are controlled
  by devres, you won't have to worry about messing up the release
  sequencing.

thanks, Tejun. So, Alex and Fabio, this patch may not be suitable currently,
since many resources at both EHCI and device side are non-managed.
-- 

Best Regards,
Peter Chen