Re: [PATCH v4 1/5] gpiolib: cdev: relocate debounce_period_us from struct gpio_desc

[Kent Gibson <warthog618@xxxxxxxxx>]

On Mon, Dec 18, 2023 at 05:01:32PM +0100, Bartosz Golaszewski wrote:
> On Mon, Dec 18, 2023 at 4:40 PM Kent Gibson <warthog618@xxxxxxxxx> wrote:
> >
> > On Mon, Dec 18, 2023 at 04:24:50PM +0100, Bartosz Golaszewski wrote:
> > > On Sat, Dec 16, 2023 at 1:17 AM Kent Gibson <warthog618@xxxxxxxxx> wrote:
> > > >

> > > > +static void line_set_debounce_period(struct line *line,
> > > > +                                    unsigned int debounce_period_us)
> > > > +{
> > > > +       bool was_suppl = line_is_supplemental(line);
> > > > +
> > > > +       WRITE_ONCE(line->debounce_period_us, debounce_period_us);
> > > > +
> > > > +       if (line_is_supplemental(line) == was_suppl)
> > > > +               return;
> > > > +
> > > > +       if (was_suppl)
> > > > +               supinfo_erase(line);
> > > > +       else
> > > > +               supinfo_insert(line);
> > >
> > > Could you add a comment here saying it's called with the config mutex
> > > taken as at first glance it looks racy but actually isn't?
> > >
> >
> > Sure.  Though it is also covered by the gdev->sem you added, right?
> > So the config_mutex is now redundant?
> > Should I document it is covered by both?
> > Or drop the config_mutex entirely?
> >
>
> No! The semaphore is a read-write semaphore and we can have multiple
> readers at once. This semaphore only protects us from the chip being
> set to NULL during a system call. It will also go away once I get the
> descriptor access serialized (or not - I'm figuring out a lockless
> approach) and finally use SRCU to protect gdev->chip.
>

Ah, so it is.

> > And you wanted some comments to explain the logic?
> > I thought this is a common "has it changed" pattern, and so didn't
> > require additional explanation, but I guess not as common as I thought.
> >
>
> If line_set_debounce_period() could be called concurrently for the
> same line, this approach would be racy. It cannot but I want a comment
> here as I fear that if in the future we add some more attributes that
> constitute "supplemental info" and which may be changed outside of the
> config lock then this would in fact become racy.
>

If any line config is going to be changed from the userspace side then
it will be by the SET_CONFIG ioctl, and so be covered by the config_mutex,
but it can't hurt to explicitly document it here as well.

Cheers,
Kent.