On Wed, Jun 27, 2012 at 1:21 PM, Minchan Kim wrote:
> When hotplug offlining happens on zone A, it starts to mark freed page
> as MIGRATE_ISOLATE type in buddy for preventing further allocation.
> (MIGRATE_ISOLATE is very irony type because it's apparently on buddy
> but we can't allocate them).
> When the memory shortage happens during hotplug offlining,
> current task starts to reclaim, then wake up kswapd.
> Kswapd checks watermark, then go sleep because current zone_watermark_ok_safe
> doesn't consider MIGRATE_ISOLATE freed page count.
> Current task continue to reclaim in direct reclaim path without kswapd's
> helping.
> The problem is that zone->all_unreclaimable is set by only kswapd
> so that current task would be looping forever like below.
>
> __alloc_pages_slowpath
> restart:
> wake_all_kswapd
> rebalance:
> __alloc_pages_direct_reclaim
> do_try_to_free_pages
> if global_reclaim && !all_unreclaimable
> return 1; /* It means we did
> did_some_progress */
> skip __alloc_pages_may_oom
> should_alloc_retry
> goto rebalance;
>
> If we apply KOSAKI's patch[1] which doesn't depends on kswapd
> about setting zone->all_unreclaimable, we can solve this problem
> by killing some task in direct reclaim path. But it doesn't wake up kswapd,
> still.
> It could be a problem still if other subsystem needs GFP_ATOMIC request.
> So kswapd should consider MIGRATE_ISOLATE when it calculate free pages
> BEFORE going sleep.
>
> This patch counts the number of MIGRATE_ISOLATE page block and
> zone_watermark_ok_safe will consider it if the system has such blocks
> (fortunately, it's very rare so no problem in POV overhead and kswapd is never
> hotpath).
>
> Copy/modify from Mel's quote
> "
> Ideal solution would be "allocating" the pageblock.
> It would keep the free space accounting as it is but historically,
> memory hotplug didn't allocate pages because it would be difficult to
> detect if a pageblock was isolated or if part of some balloon.
> Allocating just full pageblocks would work around this, However,
> it would play very badly with CMA.
> "
>
> [1] http://lkml.org/lkml/2012/6/14/74
>
> * from v1
> - add changelog
> - make functions simple
> - remove atomic variable
> - discard exact isolated free page accounting.
> - rebased on next-20120626
>
> Suggested-by: KOSAKI Motohiro
> Cc: KAMEZAWA Hiroyuki
> Cc: Aaditya Kumar
> Cc: Mel Gorman
> Signed-off-by: Minchan Kim
> ---
>
> Aaditya, coul you confirm this patch solve your problem and
> make sure nr_pageblock_isolate is zero after hotplug end?
I am really sorry for the delay.
I just tried this patch on my ARM setup.
>
> +#ifdef CONFIG_MEMORY_ISOLATION
> +static inline unsigned long nr_zone_isolate_freepages(struct zone *zone)
> +{
> + unsigned long nr_pages = 0;
> +
> + if (unlikely(zone->nr_pageblock_isolate)) {
> + nr_pages = zone->nr_pageblock_isolate * pageblock_nr_pages;
> + }
> + return nr_pages;
> +}
> +#else
> +static inline unsigned long nr_zone_isolate_freepages(struct zone *zone)
> +{
> + return 0;
> +}
> +#endif
> +
> bool zone_watermark_ok(struct zone *z, int order, unsigned long mark,
> int classzone_idx, int alloc_flags)
> {
> @@ -1629,6 +1651,14 @@ bool zone_watermark_ok_safe(struct zone *z, int order,
> unsigned long mark,
> if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
> free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
>
> + /*
> +* If the zone has MIGRATE_ISOLATE type free page,
> +* we should consider it. nr_zone_isolate_freepages is never
> +* accurate so kswapd might not sleep although she can.
> +* But it's more desirable for memory hotplug rather than
> +* forever sleep which cause livelock in direct reclaim path.
> +*/
> + free_pages -= nr_zone_isolate_freepages(z);
> return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
> free_pages);
For my test case, pages to be off lined span the whole node.
With this setup the free_pages become negative. (As you and Kamezawa-san
already expected.)
BUT because of free_pages going negative the memory off lining still livelocks
as __zone_watermark_ok() returns true.
This is because in below if comparison, because of an unsigned value
(z->lowmem_reserve[classzone_idx])
all the longs are
