Re: [PATCH v2 1/2] tmpfs: Quick token library to allow scalable retrieval of tokens from token jar

[Andrew Morton <akpm@linux-foundation.org>]

On Wed, 02 Jun 2010 10:58:42 +0200
Andi Kleen <andi@firstfloor.org> wrote:

> Andrew Morton <akpm@linux-foundation.org> writes:
> 
> I am to blame for the "token jar" name.
> 
> > OK, thanks.  But I'm still struggling a bit in understanding the
> > applicability.  Where else might one use this?  What particular
> > problem is it good at solving?
> 
> I wrote a similar scheme a long time ago for IP protocol ID
> allocation (but that code never ended up in mainline).
> Back then it was called "cookie jar" and you can actually
> still google for it :)
> 
> It can be used for pretty much anything where you have
> a global resource and want to hand it out to different CPUs,
> but still have a global limit that is enforced.
> 
> For example a file system could use it for accounting 
> free space too.
> 
> In principle you could even use it for pids for example
> or other IDs.

You could, execept the code's basically identical to percpu_counters,
only worse.

> >
> > I think the problem here is that you're using the term "token jar" as
> > if others already know what a token jar _is_.  I guess I was asleep
> > during that compsci lecture, but google doesn't appear to know about
> > the term either.
> >
> > And from looking at the tmpfs caller, it appears that the token jar is
> > being used to speed up the access to a single `unsigned long
> > free_blocks'.  Could we have used say a percpu_counter instead?
> 
> You need some synchronization, otherwise the accounting
> would not be exact and you could overflow. Yes you could
> open code it, but having it in a library is nicer.

The code doesn't have synchronisation!  qtoken_return() can modify the
per-cpu "cache" in parallel with qtoken_avail()'s walk across the
per-cpu "caches", yielding an inaccurate result.

This is all the same as percpu_add() executing in parallel with
percpu_counter_sum() or percpu_counter_sum_positive().

If we cannot tolerate that inaccuracy then these patches are no good
and we need a rethink.

If we _can_ tolerate that inaccuracy then percpu_counters can be used
here.  And doing that is preferable to reinventing percpu_counters
badly.

I'm just not seeing it.