Re: [PATCH] ACPI / GED: unregister interrupts during shutdown

[Rafael J. Wysocki]

On Thursday, December 7, 2017 9:29:58 AM CET Greg Kroah-Hartman wrote:
> On Thu, Dec 07, 2017 at 12:19:25AM +0100, Rafael J. Wysocki wrote:
> > Hi Greg,
> > 
> > On Wed, Dec 6, 2017 at 6:16 PM, Rafael J. Wysocki <rafael@xxxxxxxxxx> wrote:
> > > On Wed, Dec 6, 2017 at 5:55 PM, Sinan Kaya <okaya@xxxxxxxxxxxxxx> wrote:
> > >> On 12/6/2017 11:41 AM, Rafael J. Wysocki wrote:
> > >>> On Wed, Dec 6, 2017 at 5:11 PM, Sinan Kaya <okaya@xxxxxxxxxxxxxx> wrote:
> > >>>> On 12/6/2017 9:57 AM, Sinan Kaya wrote:
> > >>>>>> Yes, it should, so I'm not sure why you need the list in the first place.
> > >>>>>>
> > >>>>>> Also it looks like something along the lines of devres_release_all()
> > >>>>>> should be sufficient.
> > >>>>> Good suggestion, let me test this.
> > >>>>>
> > >>>>
> > >>>> I tried to pull the code into GED but the API is not public. I also looked
> > >>>> at how it is used. I was afraid to mess up with the internals of base.c by
> > >>>> calling devres_release_all() externally first and by the driver framework
> > >>>> next. I moved away from this approach.
> > >>>
> > >>> Are you sure it is called by the core in the shutdown path?
> > >>
> > >> Sorry, I was thinking about a general case not the shutdown path. If we open
> > >> this API and have device drivers call it from arbitrary places; then we could
> > >> be opening a new can of worms that show up during device driver removal.
> > 
> > [cut]
> > 
> > >
> > > OK
> > >
> > > Anyway, it looks like something is missing in the core.
> > >
> > > You shouldn't really need to do all that dance in a driver.
> > 
> > We have a problem with the ACPI GED driver which essentially is a
> > platform driver requesting a number of interrupts and handling them by
> > dispatching a specific AML method.
> > 
> > It uses devm_request_threaded_irq() to request the interrupts, so it
> > doesn't need a ->remove() callback, but it turns out to need a
> > ->shutdown() one to free all of these interrupts at the shutdown time.
> > 
> > While we can add a ->shutdown() callback to it, that essentially needs
> > to duplicate devres_release_all() somewhat.
> > 
> > Any suggestions what to do with that?
> 
> Just don't use devm_request_threaded_irq()?  :)
> 
> Seriously, those are just "helper" functions if your code happens to
> follow the pattern they provide, if not, then don't use them, it's not
> that hard to provide the correct code to unwind things properly by "open
> coding" this logic as needed.
> 
> The devm_*irq() functions are known for not being able to be used all of
> the time for lots of shutdown and cleanup issues, this isn't the first
> time it has happened, which is why we are very careful when taking
> "cleanup" patches that use those functions.

I see, thanks for the clarification.

OK, we'll need to rework the driver somewhat, then.

Thanks,
Rafael