Re: [PATCH 4/4] iio: light: apds9960: convert to use maple tree register cache

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

On Mon, Oct 27, 2025 at 01:38:06PM +0000, Jonathan Cameron wrote:
> On Mon, 27 Oct 2025 10:30:54 +0200
> Andy Shevchenko <andriy.shevchenko@intel.com> wrote:
> > On Sat, Oct 25, 2025 at 11:36:45AM +0800, Gary Chu(楚光庆) wrote:
> > > >On Fri, Oct 24, 2025 at 03:38:23PM +0800, Chu Guangqing wrote:  
> > > >> The maple tree register cache is based on a much more modern data structure
> > > >> than the rbtree cache and makes optimisation choices which are probably
> > > >> more appropriate for modern systems than those made by the rbtree cache.  

...

> > > >>   .reg_defaults = apds9960_reg_defaults,
> > > >>   .num_reg_defaults = ARRAY_SIZE(apds9960_reg_defaults),  
> > > >
> > > >^^^^ Be careful with such cases, the cache implementations may behave
> > > >differently. Have you tested this on the actual HW?
> > > >  
> > > We have conducted tests on some hardware, and performance improvements were observed,
> > >  though tests have not been carried out on all hardware models.
> > > Neither rbtree nor maple tree directly depends on hardware types (such as CPU or peripheral
> > >  models). Instead, they rely on the address distribution characteristics (discrete/continuous)
> > >  of hardware registers. The optimal cache type is determined by the hardware layout.
> > > Red-black trees excel at individual operations on discrete addresses, while Maple Trees are
> > >  proficient in range operations on contiguous addresses.  
> > 
> > It's not about the low-level cache implementation, it's about regmap
> > abstraction implementation that might differ from cache to cache
> > implementations. This all in regard how the cold cache is getting filled up.
> > There is a separate discussion (unrelated to the topic of your series) where
> > this was brought up.
> 
> I appreciate these things can be hard to track down with lots of threads in flight
> but any chance of a reference for that? I'd be a little surprised if these uses
> are complicated enough to hit corner cases but would like to know more.
> I've taken a few similar changes in the past thinking there would be no
> practical difference.

Sure, it appeared in the discussion of v2 of the following patch:
https://lore.kernel.org/linux-gpio/20251009132651.649099-2-bigunclemax@gmail.com/

> > That's why I asked how this was tested.
> > 
> > In any case, up to Jonathan, but I had to rise a potential misbehave, so in my
> > opinion this kind of corner cases needs to be tested on real HW.
> > 
> > > >>   .max_register = APDS9960_REG_GFIFO_DIR(RIGHT),
> > > >> - .cache_type = REGCACHE_RBTREE,
> > > >> + .cache_type = REGCACHE_MAPLE,
> > > >>  };  

-- 
With Best Regards,
Andy Shevchenko