Re: [PATCH 2/2] zsmalloc: fine-grained inuse ratio based fullness grouping

From: Yosry Ahmed
Date: Fri Feb 10 2023 - 01:08:22 EST


On Thu, Feb 9, 2023 at 10:04 PM Yosry Ahmed <yosryahmed@xxxxxxxxxx> wrote:
>
> On Thu, Feb 9, 2023 at 9:53 PM Sergey Senozhatsky
> <senozhatsky@xxxxxxxxxxxx> wrote:
> >
> > On (23/02/09 21:15), Yosry Ahmed wrote:
> > > > We need to be able to do zs_stat_get(class, CLASS_USAGE_70) or
> > > > zs_stat_get(class, CLASS_USAGE_10) in zs_stats_size_show() to
> > > > show class's fullness stats.
> > >
> > > If we use #define FULLNESS_GROUPS 10 for example, we can break down
> > > struct zs_size_stat from a single array to two arrays, one of the for
> > > fullness groups and the other one for the rest of the stats (e.g.
> > > OBJ_USED). We can have different helpers to update each, the former
> > > taking in a fullness value (0 to FULLNESS_GROUPS-1), and the latter an
> > > enum. WDYT?
> >
> > Wouldn't this be more code to review, maintain though? I mean,
> > what do we gain with this - fewer values in enum?
>
> I don't feel very strongly, but the patch above has a lot of hardcoded
> values, multiple enums & static arrays. If we want to change the
> number of fullness groups to 5 or 20, we modify all of them. Just
> trying to make the code cleaner. If you think it will add a lot more
> code to review/maintain, then maybe it's not worth it. I didn't really
> give it a lot of thought.

We can also do something like:

#define NR_FULLNESS_GROUPS 10

enum class_stat_type {
OBJ_ALLOCATED = NR_FULLNESS_GROUPS,
OBJ_USED,
NR_ZS_STAT_TYPE,
};

and get rid of enum fullness_group, I think the current stats inc/dec
helpers will work out of the box.