Re: [PATCH 03/10] mm/page_alloc: handle page on pcp correctly if it's pageblock is isolated
From: Joonsoo Kim
Date: Mon Jul 14 2014 - 02:18:37 EST
On Mon, Jul 07, 2014 at 05:19:48PM +0200, Vlastimil Babka wrote:
> On 07/04/2014 09:57 AM, Joonsoo Kim wrote:
> >If pageblock of page on pcp are isolated now, we should free it to isolate
> >buddy list to prevent future allocation on it. But current code doesn't
> >do this.
> >Moreover, there is a freepage counting problem on current code. Although
> >pageblock of page on pcp are isolated now, it could go normal buddy list,
> >because get_onpcp_migratetype() will return non-isolate migratetype.
> get_onpcp_migratetype() is only introduced in later patch.
Yes, I will fix it.
> >In this case, we should do either adding freepage count or changing
> >migratetype to MIGRATE_ISOLATE, but, current code do neither.
> I wouldn't say it "do neither". It already limits the freepage
> counting to !MIGRATE_ISOLATE case (and it's not converted to
> __mod_zone_freepage_state for some reason). So there's accounting
> mismatch in addition to buddy list misplacement.
> >This patch fixes these two problems by handling pageblock migratetype
> >before calling __free_one_page(). And, if we find the page on isolated
> >pageblock, change migratetype to MIGRATE_ISOLATE to prevent future
> >allocation of this page and freepage counting problem.
> So although this is not an addition of a new pageblock migratetype
> check to the fast path (the check is already there), I would prefer
> removing the check :)
Yes, I want to do it if possible. :)
> With the approach of pcplists draining
> outlined in my reply to 00/10, we would allow a misplacement to
> happen (and the page accounted as freepage) immediately followed by
> move_frepages_block which would place the page onto isolate freelist
> with the rest. Anything newly freed will get isolate_migratetype
> determined in free_hot_cold_page or __free_pages_ok (where it would
> need moving the migratepage check under the disabled irq part) and
> be placed and buddy-merged properly.
I explained the problem of this approach in 00/10.
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/