Re: [PATCH 3/3] mm: when handling percpu_pagelist_fraction, use on_each_cpu()to set percpu pageset fields.

[Cody P Schafer]

On 04/08/2013 10:28 AM, Cody P Schafer wrote:
> On 04/08/2013 05:20 AM, Gilad Ben-Yossef wrote:
> > On Fri, Apr 5, 2013 at 11:33 PM, Cody P Schafer
> > <cody@xxxxxxxxxxxxxxxxxx> wrote:
> > > In free_hot_cold_page(), we rely on pcp->batch remaining stable.
> > > Updating it without being on the cpu owning the percpu pageset
> > > potentially destroys this stability.
> > >
> > > Change for_each_cpu() to on_each_cpu() to fix.
> >
> > Are you referring to this? -
>
> This was the case I noticed.
>
> > 1329         if (pcp->count >= pcp->high) {
> > 1330                 free_pcppages_bulk(zone, pcp->batch, pcp);
> > 1331                 pcp->count -= pcp->batch;
> > 1332         }
> >
> > I'm probably missing the obvious but won't it be simpler to do this in
> >   free_hot_cold_page() -
> >
> > 1329         if (pcp->count >= pcp->high) {
> > 1330                  unsigned int batch = ACCESS_ONCE(pcp->batch);
> > 1331                 free_pcppages_bulk(zone, batch, pcp);
> > 1332                 pcp->count -= batch;
> > 1333         }
>
> Potentially, yes. Note that this was simply the one case I noticed,
> rather than certainly the only case.
>
> I also wonder whether there could be unexpected interactions between
> ->high and ->batch not changing together atomically. For example, could
> adjusting this knob cause ->batch to rise enough that it is greater than
> the previous ->high? If the code above then runs with the previous
> ->high, ->count wouldn't be correct (checking this inside
> free_pcppages_bulk() might help on this one issue).
>
> > Now the batch value used is stable and you don't have to IPI every CPU
> > in the system just to change a config knob...
>
> Is this really considered an issue? I wouldn't have expected someone to
> adjust the config knob often enough (or even more than once) to cause
> problems. Of course as a "It'd be nice" thing, I completely agree.

Would using schedule_on_each_cpu() instead of on_each_cpu() be an 
improvement, in your opinion?

--
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/