Re: [PATCH 0/5 v3] irq / PM: Suspend-to-idle wakeup interrupts

["Rafael J. Wysocki" <rjw@rjwysocki.net>]

On Friday, August 29, 2014 12:44:11 AM Thomas Gleixner wrote:
> On Wed, 27 Aug 2014, Rafael J. Wysocki wrote:
> > To me, all of this is relatively straightforward and the handling of
> > IRQF_NO_SUSPEND for shared interrupts, which is a separate problem, can be
> > addressed on top of it later (make no mistake, I still think that it should be
> > addressed).
> 
> Why? Just because there is enough idiotic code using it?
> 
> Let's look at the usage sites:
> 
> sound/soc/codecs/twl6040.c
> 
>   Introduced in commit 8d61f4901f83461e1f04df7743777e9db5f541aa
> 
>     ASoC: twl6040: Convert PLUGINT to no-suspend irq
>     
>     Convert headset PLUGINT interrupt to NO_SUSPEND type in order to
>     allow handling of insertion/removal events while device is
>     suspended.
> 
>   So why does this need to be a NO_SUSPEND type interrupt? Just because
>   the flag is sexy? What's wrong with using the wake mechanism?
> 
> drivers/xen/events/events_base.c
> 
>   A genuine usecase which makes sense and is not shared
> 
> drivers/watchdog/intel-mid_wdt.c
> 
>   MUHAHAHAHAHAHA
> 
>   static irqreturn_t mid_wdt_irq(int irq, void *dev_id)
>   {
>       panic("Kernel Watchdog");
>       /* This code should not be reached */
>       return IRQ_HANDLED;
>   }
> 
>   So why does this need to be NO_SUSPEND? Because we forgot to trigger
>   the watchdog during suspend? Brilliant!
> 
> drivers/usb/phy/phy-ab8500-usb.c
> 
>   Of course no explanation WHY this uses NO_SUSPEND plus SHARED and
>   there is no fcking reason to do so.
> 
> So really, I'm too lazy to walk through that mess further. I bet NONE
> of the usage sites except those for which this has been introduced in
> the first place has a real good reason to do so.

A quite similar conclusion occured to me earlier today independently,
so we seem to be in agreement here. :-)

> Unless someone comes up with a use case where a shared NO_SUSPEND
> handler is absolutely required, there is nothing which needs to be
> addressed. You've proven yourself, that the wake mechanism is
> sufficient to solve the problem which led to this discussion.

Yeah, that particular thing can be perfectly addressed without using
IRQF_NO_SUSPEND.

There seems to be quite some confusion about that in the ARM world,
though, as IRQF_NO_SUSPEND things *happen* to wake the system up if WFI is
used to emulate "suspend" and that appears to make people believe that using
them for this purpose is actually fine (because that appears to work).

> So we rather go and fix the ABUSE instead of making it legitimate by
> fugly workarounds in the core code.

Fine by me.

> The same applies to the IRQF_RESUME_EARLY flag.
> 
> Just look at:
> 
>   commit 8b41669ceba0c2d4c09d69ccb9a3458953dae784
> 
>     mfd: twl4030: Fix chained irq handling on resume from suspend
>     
>     The irqs are enabled one-by-one in pm core resume_noirq phase.
>     This leads to situation where the twl4030 primary interrupt
>     handler (PIH) is enabled before the chained secondary handlers
>     (SIH). As the PIH cannot clear the pending interrupt, and
>     SIHs have not been enabled yet, a flood of interrupts hangs
>     the device.
>     
>     Fixed the issue by setting the SIH irqs with IRQF_EARLY_RESUME
>     flags, so they get enabled before the PIH.
> 
> So we solve an ordering problem which has a completely different root
> cause by slapping random flags on it which paper over the issue?
> 
> The solution to the problem is completely wrong and of course the
> "fix" is only applied to the one instance which might be affected by
> that issue, i.e. the one which caused the patch submitter trouble.
> 
> Now I don't blame the author, I blame the maintainer who happily
> applied that "fix".
> 
> That's unfortunately a very common pattern in the kernel which will
> cause serious maintainability issues in the long run, but of course
> that's just the opinion of a grumpy old greybeard.
> 
> The sad thing is that neither the author nor the maintainer who
> applied it is going to be around and responsive when the shit hits the
> fan.
>
> That commit is a perfect example for this.

Oh, I can reach that particular maintainer if need be.

That said I was not involved in the introduction of IRQF_EARLY_RESUME (or
at least I can't recall being involved) and I also can't recall having used
it for anything, so I can't really say what the original motivation was.

Rafael