Re: [PATCH 02/35] writeback: safety margin for bdi stat error

[Wu Fengguang]

On Thu, Jan 13, 2011 at 05:59:49AM +0800, Jan Kara wrote:
> On Mon 13-12-10 22:46:48, Wu Fengguang wrote:
> > In a simple dd test on a 8p system with "mem=256M", I find all light
> > dirtier tasks on the root fs are get heavily throttled. That happens
> > because the global limit is exceeded. It's unbelievable at first sight,
> > because the test fs doing the heavy dd is under its bdi limit.  After
> > doing some tracing, it's discovered that
> > 
> >         bdi_dirty < bdi_dirty_limit() < global_dirty_limit() < nr_dirty
>           ^^ bdi_dirty is the number of pages dirtied on BDI? I.e.
> bdi_nr_reclaimable + bdi_nr_writeback?

Yes.

> > So the root cause is, the bdi_dirty is well under the global nr_dirty
> > due to accounting errors. This can be fixed by using bdi_stat_sum(),
>   So which statistic had the big error? I'd just like to understand
> this (and how come your patch improves the situation)...

bdi_stat_error() = nr_cpu_ids * BDI_STAT_BATCH
                 = 8 * (8*(1+ilog2(8)))
                 = 8 * 8 * 4
                 = 256 pages
                 = 1MB

> > however that's costly on large NUMA machines. So do a less costly fix
> > of lowering the bdi limit, so that the accounting errors won't lead to
> > the absurd situation "global limit exceeded but bdi limit not exceeded".
> > 
> > This provides guarantee when there is only 1 heavily dirtied bdi, and
> > works by opportunity for 2+ heavy dirtied bdi's (hopefully they won't
> > reach big error _and_ exceed their bdi limit at the same time).
> > 
> ...
> > @@ -458,6 +464,14 @@ unsigned long bdi_dirty_limit(struct bac
> >  	long numerator, denominator;
> >  
> >  	/*
> > +	 * try to prevent "global limit exceeded but bdi limit not exceeded"
> > +	 */
> > +	if (likely(dirty > bdi_stat_error(bdi)))
> > +		dirty -= bdi_stat_error(bdi);
> > +	else
> > +		return 0;
> > +
>   Ugh, so if by any chance global_dirty_limit() <= bdi_stat_error(bdi), you
> will limit number of unreclaimable pages for that bdi 0? Why?

Good catch! Yeah it may lead to regressions and should be voided.

Thanks,
Fengguang
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/