Re: [PATCH 01/11] mm: page_alloc: set_migratetype_isolate: drain PCP prior to isolating

["Michal Nazarewicz" <mina86@mina86.com>]

> On Thu, Dec 29, 2011 at 2:39 PM, Marek Szyprowski
> <m.szyprowski@samsung.com> wrote:
>> From: Michal Nazarewicz <mina86@mina86.com>
>>
>> When set_migratetype_isolate() sets pageblock's migrate type, it does
>> not change each page_private data.  This makes sense, as the function
>> has no way of knowing what kind of information page_private stores.

>> A side effect is that instead of draining pages from all zones,
>> set_migratetype_isolate() now drain only pages from zone pageblock it
>> operates on is in.

>> +/* Caller must hold zone->lock. */
>> +static void __zone_drain_local_pages(void *arg)
>> +{
>> +       struct per_cpu_pages *pcp;
>> +       struct zone *zone = arg;
>> +       unsigned long flags;
>> +
>> +       local_irq_save(flags);
>> +       pcp = &per_cpu_ptr(zone->pageset, smp_processor_id())->pcp;
>> +       if (pcp->count) {
>> +               /* Caller holds zone->lock, no need to grab it. */
>> +               __free_pcppages_bulk(zone, pcp->count, pcp);
>> +               pcp->count = 0;
>> +       }
>> +       local_irq_restore(flags);
>> +}
>> +
>> +/*
>> + * Like drain_all_pages() but operates on a single zone.  Caller must
>> + * hold zone->lock.
>> + */
>> +static void __zone_drain_all_pages(struct zone *zone)
>> +{
>> +       on_each_cpu(__zone_drain_local_pages, zone, 1);
>> +}
>> +

On Sun, 01 Jan 2012 08:49:13 +0100, Gilad Ben-Yossef <gilad@benyossef.com> wrote:
> Please consider whether sending an IPI to all processors in the system
> and interrupting them is appropriate here.
>
> You seem to assume that it is probable that each CPU of the possibly
> 4,096 (MAXSMP on x86) has a per-cpu page for the specified zone,

I'm not really assuming that (in fact I expect what you fear, ie. that
most CPUs won't have pages from specified zone an PCP list), however,
I really need to make sure to get them off all PCP lists.

> otherwise you're just interrupting them out of doing something useful,
> or save power idle for nothing.

Exactly what's happening now anyway.

> While that may or may not be a reasonable assumption for the general
> drain_all_pages that drains pcps from all zones, I feel it is less
> likely to be the right thing once you limit the drain to a single
> zone.

Currently, set_migratetype_isolate() seem to do more then it needs to,
ie. it drains all the pages even though all it cares about is a single
zone.

> Some background on my attempt to reduce "IPI noise" in the system in
> this context is probably useful here as
> well: https://lkml.org/lkml/2011/11/22/133

Looks interesting, I'm not entirely sure why it does not end up a race
condition, but in case of __zone_drain_all_pages() we already hold
zone->lock, so my fears are somehow gone..  I'll give it a try, and prepare
a patch for __zone_drain_all_pages().

-- 
Best regards,                                         _     _
.o. | Liege of Serenely Enlightened Majesty of      o' \,=./ `o
..o | Computer Science,  Michał “mina86” Nazarewicz    (o o)
ooo +----<email/xmpp: mpn@google.com>--------------ooO--(_)--Ooo--