Re: [PATCH v2 5/6] alloc_tag: make page allocation tag reference size configurable

[Kent Overstreet]

On Tue, Sep 03, 2024 at 06:07:28PM GMT, Suren Baghdasaryan wrote:
> On Sun, Sep 1, 2024 at 10:09 PM Andrew Morton <akpm@xxxxxxxxxxxxxxxxxxxx> wrote:
> >
> > On Sun,  1 Sep 2024 21:41:27 -0700 Suren Baghdasaryan <surenb@xxxxxxxxxx> wrote:
> >
> > > Introduce CONFIG_PGALLOC_TAG_REF_BITS to control the size of the
> > > page allocation tag references. When the size is configured to be
> > > less than a direct pointer, the tags are searched using an index
> > > stored as the tag reference.
> > >
> > > ...
> > >
> > > +config PGALLOC_TAG_REF_BITS
> > > +     int "Number of bits for page allocation tag reference (10-64)"
> > > +     range 10 64
> > > +     default "64"
> > > +     depends on MEM_ALLOC_PROFILING
> > > +     help
> > > +       Number of bits used to encode a page allocation tag reference.
> > > +
> > > +       Smaller number results in less memory overhead but limits the number of
> > > +       allocations which can be tagged (including allocations from modules).
> > > +
> >
> > In other words, "we have no idea what's best for you, you're on your
> > own".
> >
> > I pity our poor users.
> >
> > Can we at least tell them what they should look at to determine whether
> > whatever random number they chose was helpful or harmful?
> 
> At the end of my reply in
> https://lore.kernel.org/all/CAJuCfpGNYgx0GW4suHRzmxVH28RGRnFBvFC6WO+F8BD4HDqxXA@xxxxxxxxxxxxxx/#t
> I suggested using all unused page flags. That would simplify things
> for the user at the expense of potentially using more memory than we
> need.

Why would that use more memory, and how much?

> In practice 13 bits should be more than enough to cover all
> kernel page allocations with enough headroom for page allocations
> coming from loadable modules. I guess using 13 as the default would
> cover most cases. In the unlikely case a specific system needs more
> tags, the user can increase this value. It can also be set to 64 to
> force direct references instead of indexing for better performance.
> Would that approach be acceptable?

Any knob that has to be kept track of and adjusted is a real hassle -
e.g. lockdep has a bunch of knobs that have to be periodically tweaked,
that's used by _developers_, and they're often wrong.