Re: [PATCH] gpio: use raw spinlock for gpio chip shadowed data

[Bartosz Golaszewski <brgl@xxxxxxxx>]

On Mon, Apr 18, 2022 at 5:43 PM Schspa Shi <schspa@xxxxxxxxx> wrote:
>
> Andy Shevchenko <andy.shevchenko@xxxxxxxxx> writes:
>
> > On Mon, Apr 18, 2022 at 2:38 PM Andy Shevchenko
> > <andy.shevchenko@xxxxxxxxx> wrote:
> >> On Mon, Apr 18, 2022 at 6:07 AM Schspa Shi <schspa@xxxxxxxxx> wrote:
> >> > Andy Shevchenko <andy.shevchenko@xxxxxxxxx> writes:
> >>
> >> ...
> >>
> >> > >   drivers/gpio/gpio-mmio.c    | 22 +++++++++++-----------
> >> > >   include/linux/gpio/driver.h |  2 +-
> >> > >
> >> > > You can’t do it for one driver only. As I told it will require too much of additional churn to make this to be series.
> >> > >
> >> >
> >> > It seems I have misunderstood your "too much of additional churn". Can
> >> > you explain it?
> >> > The gpio-mmio.c and driver.h here are the basics of other gpio
> >> > drivers. In my opinion, these two files
> >> > belong to the basic code of gpio, and functions such as bgpio_init are
> >> > declared in
> >> > include/linux/gpio/driver.h and implemented in
> >> > drivers/gpio/gpio-mmio.c. So there is no churn.
> >>
> >> When you change the member of the data structure, you have to change
> >> all its users. You can't change only one at a time because it will be
> >> a (compile-time) bisectability issue.
>
>
> Yes, I understand and will take for bisectability use case for the next time.
>
> >
> > Answering your question here, it will require moving to union with an
> > additional member and corresponding core changes, convert all drivers
> > one-by-one, and remove the old type. It's not worth doing it, but as I
> > said let maintainers decide.
>
> Okay, sorry for my misunderstanding, I thought you were saying it's
> bad to modify too many different files in one patch, so I split the
> patch into a series of patchsets.
>
> So, let Linus Walleij or Bartosz Golaszewski to decide for it ?
> I have the same options as you, it's a small change, and no need to
> trouble everyone for it.
>

I prefer a single patch in this case, we can apply it closer to the
next merge window.

Bart

> Because this structure will be used as the same instance in multiple
> files, even if we change this variable to union first, it can be
> compiled, but the program will still not work properly. This is
> because bgpio_init is initialized with the type of raw_spinlock_t,
> but is still accessed as spinlock_t in other drivers, which is a
> serious abnormal initialization.