On Tue, Oct 15, 2024 at 01:31:40PM +0800, Chen-Yu Tsai wrote:
> On Mon, Oct 14, 2024 at 7:20 PM Andy Shevchenko
> <andriy.shevchenko@xxxxxxxxxxxxxxx> wrote:
> > On Mon, Oct 14, 2024 at 12:06:16PM +0800, Chen-Yu Tsai wrote:
> > > On Thu, Oct 10, 2024 at 11:20 PM Andy Shevchenko
> > > <andriy.shevchenko@xxxxxxxxxxxxxxx> wrote:
> > > > On Tue, Oct 08, 2024 at 03:34:25PM +0800, Chen-Yu Tsai wrote:
...
> > > > > +static void i2c_of_probe_simple_disable_gpio(struct device *dev, struct i2c_of_probe_simple_ctx *ctx)
> > > > > +{
> > > > > + if (!ctx->gpiod)
> > > > > + return;
> > > >
> > > > Do you need this check for the future patches?
> > >
> > > Not sure I follow. The check is needed because this function is called
> > > in i2c_of_probe_simple_cleanup(), but the GPIO could have been released
> > > earlier in i2c_of_probe_simple_cleanup_early(), and that makes this
> > > function a no-op.
> >
> > Do you have a known race condition then? This is bad. You shouldn't rely on
> > the sequence of events here, or the serialisation has to be added.
>
> No there isn't. Explanation below.
>
> > > The helpers for the release side are quite short, but the ones on the
> > > request side wrap some conditional and error handling. I think it's
> > > better to keep it symmetric?
> >
> > Yes, but why do you need the above check, I didn't still get...
> > I.o.w. you think that there is a gap in time that (if no check) the GPIO
> > descriptor might be changed? But then how does it affect anyway the possibility
> > that it becomes not NULL even with the current code.
>
> There are two codes paths, either
>
> a) successfully finding a device and enabling it, or
> b) exhausting all options and not finding a device, because it was
> optional or it is malfunctioning.
>
> After either code path, this cleanup function is called.
>
> In path (a), the GPIO descriptor is released prior to enabling the device,
> because the descriptor is an exclusive resource, and as soon as the device
> is enabled, its corresponding driver may probe and request the same GPIO,
> and would fail if it was not released.
>
> In path (b), nothing was enabled, and the GPIO descriptor was not released
> early.
>
> The cleanup function here accounts for both cases, hence the check.
Yes, but the very same check is inside gpiod_set_value(). I'm still puzzled
about the duplication. Maybe I'm missing something...
> A step-by-step description might be clearer:
>
> 1. i2c_of_probe_simple_enable()
> ...
> 1a. i2c_of_probe_simple_get_supply()
> 1b. i2c_of_probe_simple_get_gpiod()
> 1c. i2c_of_probe_simple_enable_regulator()
> 1d. i2c_of_probe_simple_set_gpio()
>
> 2. Loop through potential component options and probe; if one is found:
> 2a. i2c_of_probe_simple_cleanup_early()
> 2a-i. i2c_of_probe_simple_put_gpiod
> 2b. Enable device and driver's probe() gets called
>
> 3. i2c_of_probe_simple_cleanup()
> 3a. i2c_of_probe_simple_disable_gpio()
> 3b. i2c_of_probe_simple_put_gpiod()
> 3c. i2c_of_probe_simple_disable_regulator()
> 3d. i2c_of_probe_simple_put_supply()
>
> > > > > + /* Ignore error if GPIO is not in output direction */
> > > > > + gpiod_set_value(ctx->gpiod, !ctx->opts->gpio_assert_to_enable);
> > > > > +}
--
With Best Regards,
Andy Shevchenko
