Re: [PATCH v18] pwm: airoha: Add support for EN7581 SoC

[Andy Shevchenko <andriy.shevchenko@linux.intel.com>]

On Mon, Jun 30, 2025 at 12:19:22PM +0200, Christian Marangi wrote:
> On Fri, Jun 27, 2025 at 01:56:34PM +0300, Andy Shevchenko wrote:
> > On Fri, Jun 27, 2025 at 12:34:49PM +0200, Christian Marangi wrote:
> > > On Fri, Jun 27, 2025 at 01:25:48PM +0300, Andy Shevchenko wrote:
> > > > On Fri, Jun 27, 2025 at 11:32:46AM +0200, Christian Marangi wrote:
> > > > > On Fri, Jun 27, 2025 at 11:58:04AM +0300, Andy Shevchenko wrote:
> > > > > > On Fri, Jun 27, 2025 at 12:47:53AM +0200, Christian Marangi wrote:

...

> > > > > > > +	/* Global mutex to protect bucket used refcount_t */
> > > > > > > +	struct mutex mutex;
> > > > > > 
> > > > > > This makes a little sense. Either you use refcount_t (which is atomic) or
> > > > > > use mutex + regular variable.
> > > > > 
> > > > > Using a regular variable I lose all the benefits of refcount_t with
> > > > > underflow and other checks.
> > > > 
> > > > Then drop the mutex, atomic operations do not need an additional
> > > > synchronisation. Btw, have you looked at kref APIs? Maybe that
> > > > would make the intention clearer?
> > > 
> > > It's needed for
> > > 
> > > +       mutex_lock(&pc->mutex);
> > > +       if (refcount_read(&pc->buckets[bucket].used) == 0) {
> > > +               config_bucket = true;
> > > +               refcount_set(&pc->buckets[bucket].used, 1);
> > > +       } else {
> > > +               refcount_inc(&pc->buckets[bucket].used);
> > > +       }
> > > +       mutex_unlock(&pc->mutex);
> > > 
> > > the refcount_read + refcount_set.
> > 
> > Which is simply wrong. Nobody should use atomics in such a way.
> > Imagine if somebody wants to copy something like this in their
> > code (in case of no mutex is there), they most likely won't notice
> > this subtle bug.
> >
> 
> Yes I understand that someone might think the additional mutex can be
> ""optional""
> 
> > > As you explained there might be case where refcount_read is zero but nother
> > > PWM channel is setting the value so one refcount gets lost.
> > 
> > Right, because you should use refcount_inc_and_test() and initialise it
> > to -MAX instead of 0. Or something like this.
> 
> Mhhh I think API for _inc_and_test doesn't currently exist and I don't
> feel too confident implementing them currently.

Ther is refcount_inc_not_zero(), but the main point here that refcount_t seems
not fit the case. It doesn't work with negative values, and 0 is special.

> > > kref I checked but not useful for the task.
> > 
> > Okay.
> > 
> > > The logic here is
> > > 
> > > - refcount init as 0 (bucket unused)
> > > - refcount set to 1 on first bucket use (bucket get configured)
> > > - refcount increased if already used
> > > - refcount decreased when PWM channel released
> > > - bucket gets flagged as unused when refcount goes to 0 again
> 
> Do you think I should bite the bullet and just drop using refcount and
> implement a simple int variable protected by a mutex?

Yes.

-- 
With Best Regards,
Andy Shevchenko