Re: [PATCH 0/5] gpio: cdev: bail out of poll() if the device goes down

[Kent Gibson]

On Thu, Aug 17, 2023 at 09:27:37AM +0200, Bartosz Golaszewski wrote:
> On Thu, Aug 17, 2023 at 6:41 AM Kent Gibson <warthog618@xxxxxxxxx> wrote:
> >
> > On Wed, Aug 16, 2023 at 11:41:06PM +0200, Linus Walleij wrote:
> > > On Wed, Aug 16, 2023 at 2:20 PM Bartosz Golaszewski <brgl@xxxxxxxx> wrote:
> > >
> > > > From: Bartosz Golaszewski <bartosz.golaszewski@xxxxxxxxxx>
> > > >
> > > > Wake up all three wake queues (the one associated with the character
> > > > device file, the one for V1 line events and the V2 line request one)
> > > > when the underlying GPIO device is unregistered. This way we won't get
> > > > stuck in poll() after the chip is gone as user-space will be forced to
> > > > go back into a new system call and will see that gdev->chip is NULL.
> > > >
> > > > Bartosz Golaszewski (5):
> > > >   gpio: cdev: ignore notifications other than line status changes
> > > >   gpio: cdev: rename the notifier block and notify callback
> > > >   gpio: cdev: wake up chardev poll() on device unbind
> > > >   gpio: cdev: wake up linereq poll() on device unbind
> > > >   gpio: cdev: wake up lineevent poll() on device unbind
> > >
> > > I see why this is needed and while the whole notification chain
> > > is a bit clunky I really cannot think about anything better so:
> > > Reviewed-by: Linus Walleij <linus.walleij@xxxxxxxxxx>
> > >
> >
> > The issue I have is with the repurposing/reuse of the existing notifier
> > block that sends line changed events to the chardev.
> > Correct me if I'm wrong, but now all line requests will receive those
> > events as well.
> > They have no business receiving those events, and it scales badly.
> >
> > My preference would be for a separate nb for the chip removal to keep
> > those two classes of events distinct.
> >
> 
> I would normally agree if there was a risk of abuse of those
> notifications by drivers but this is all private to gpiolib. And line
> requests that receive line state notifications simply ignore them.
> This isn't a bottleneck codepath IMO so where's the issue? We would be
> using a second notifier head of 40 bytes to struct gpio_device for no
> reason.
> 

Yeah, this is a space/time trade-off, and you've gone with space over
time.  I would select time over space.
40 bytes per device is negligable, and there is never a case where the
line request wants to see a change event - it either relates to a
different request, or it was triggered by the request itself.
Is there an echo in here ;-)?

Cheers,
Kent.