Re: [PATCH] writeback: fix writeback cache thrashing

[Fengguang Wu]

On Sat, Jan 05, 2013 at 03:41:54AM -0600, Simon Jeons wrote:
> On Sat, 2013-01-05 at 15:38 +0800, Fengguang Wu wrote:
> > On Fri, Jan 04, 2013 at 11:26:43PM -0600, Simon Jeons wrote:
> > > On Sat, 2013-01-05 at 11:26 +0800, Fengguang Wu wrote:
> > > > > > > Hi Namjae,
> > > > > > >
> > > > > > > Why use bdi_stat_error here? What's the meaning of its comment "maximal
> > > > > > > error of a stat counter"?
> > > > > > Hi Simon,
> > > > > > 
> > > > > > As you know bdi stats (BDI_RECLAIMABLE, BDI_WRITEBACK â) are kept in
> > > > > > percpu counters.
> > > > > > When these percpu counters are incremented/decremented simultaneously
> > > > > > on multiple CPUs by small amount (individual cpu counter less than
> > > > > > threshold BDI_STAT_BATCH),
> > > > > > it is possible that we get approximate value (not exact value) of
> > > > > > these percpu counters.
> > > > > > In order, to handle these percpu counter error we have used
> > > > > > bdi_stat_error. bdi_stat_error is the maximum error which can happen
> > > > > > in percpu bdi stats accounting.
> > > > > > 
> > > > > > bdi_stat(bdi, BDI_RECLAIMABLE);
> > > > > >  -> This will give approximate value of BDI_RECLAIMABLE by reading
> > > > > > previous value of percpu count.
> > > > > > 
> > > > > > bdi_stat_sum(bdi, BDI_RECLAIMABLE);
> > > > > >  ->This will give exact value of BDI_RECLAIMABLE. It will take lock
> > > > > > and add current percpu count of individual CPUs.
> > > > > >    It is not recommended to use it frequently as it is expensive. We
> > > > > > can better use âbdi_statâ and work with approx value of bdi stats.
> > > > > > 
> > > > > 
> > > > > Hi Namjae, thanks for your clarify.
> > > > > 
> > > > > But why compare error stat count to bdi_bground_thresh? What's the
> > > > 
> > > > It's not comparing bdi_stat_error to bdi_bground_thresh, but rather,
> > > > in concept, comparing bdi_stat (with error bound adjustments) to
> > > > bdi_bground_thresh.
> > > > 
> > > > > relationship between them? I also see bdi_stat_error compare to
> > > > > bdi_thresh/bdi_dirty in function balance_dirty_pages. 
> > > > 
> > > 
> > > Hi Fengguang,
> > > 
> > > > Here, it's trying to use bdi_stat_sum(), the accurate (however more
> > > > costly) version of bdi_stat(), if the error would possibly be large:
> > > 
> > > Why error is large use bdi_stat_sum and error is few use bdi_stat?
> > 
> 
> Thanks for your response Fengguang! :)

You are welcome.

> > It's the opposite. Please check this per-cpu counter routine to get an idea:
> > 
> > /*
> >  * Add up all the per-cpu counts, return the result.  This is a more accurate
> >  * but much slower version of percpu_counter_read_positive()
> >  */                                                 
> > s64 __percpu_counter_sum(struct percpu_counter *fbc)
> > 
> > > > 
> > > >                 if (bdi_thresh < 2 * bdi_stat_error(bdi)) {
> > > >                         bdi_reclaimable = bdi_stat_sum(bdi, BDI_RECLAIMABLE);
> > > >                         //...
> > > >                 } else {
> > > >                         bdi_reclaimable = bdi_stat(bdi, BDI_RECLAIMABLE);
> > > >                         //...
> > > >                 }
> > > > 
> 
> The comment above these codes:
> 
>                  * In order to avoid the stacked BDI deadlock we need
>                  * to ensure we accurately count the 'dirty' pages when
>                  * the threshold is low.
> 
> Why your meaning threshold low is error large? 

Because bdi_reclaimable is normally less than or at least comparable
to bdi_thresh. So (bdi_thresh < 2 * bdi_stat_error(bdi)) means the
resulted bdi_reclaimable will be small, so small that the bdi_stat()
error is not ignorable and should be avoided.

Thanks,
Fengguang

> 
> > > > Here the comment should have explained it well:
> > > > 
> > > >                  * In theory 1 page is enough to keep the comsumer-producer
> > > >                  * pipe going: the flusher cleans 1 page => the task dirties 1
> > > >                  * more page. However bdi_dirty has accounting errors.  So use
> > > 
> > > Why bdi_dirty has accounting errors?
> > 
> > Because it typically uses bdi_stat() to get the rough sum of the per-cpu
> > counters.
> >  
> > Thanks,
> > Fengguang
> > 
> > > >                  * the larger and more IO friendly bdi_stat_error.
> > > >                  */
> > > >                 if (bdi_dirty <= bdi_stat_error(bdi))
> > > >                         break;
> > > > 
> > > > 
> > > > 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/