Re: [RFC v2] memory-hotplug: remove MIGRATE_ISOLATE fromfree_area->free_list

[Bartlomiej Zolnierkiewicz]

Hi,

On Thursday 06 September 2012 04:53:38 Minchan Kim wrote:
> Normally, MIGRATE_ISOLATE type is used for memory-hotplug.
> But it's irony type because the pages isolated would exist
> as free page in free_area->free_list[MIGRATE_ISOLATE] so people
> can think of it as allocatable pages but it is *never* allocatable.
> It ends up confusing NR_FREE_PAGES vmstat so it would be
> totally not accurate so some of place which depend on such vmstat
> could reach wrong decision by the context.
> 
> There were already report about it.[1]
> [1] 702d1a6e, memory-hotplug: fix kswapd looping forever problem
> 
> Then, there was other report which is other problem.[2]
> [2] http://www.spinics.net/lists/linux-mm/msg41251.html
> 
> I believe it can make problems in future, too.
> So I hope removing such irony type by another design.
> 
> I hope this patch solves it and let's revert [1] and doesn't need [2].
> 
> * Changelog v1
>  * Fix from Michal's many suggestion
> 
> Cc: Michal Nazarewicz <mina86@xxxxxxxxxx>
> Cc: Mel Gorman <mel@xxxxxxxxx>
> Cc: Kamezawa Hiroyuki <kamezawa.hiroyu@xxxxxxxxxxxxxx>
> Cc: Yasuaki Ishimatsu <isimatu.yasuaki@xxxxxxxxxxxxxx>
> Cc: Wen Congyang <wency@xxxxxxxxxxxxxx>
> Cc: Konrad Rzeszutek Wilk <konrad.wilk@xxxxxxxxxx>
> Signed-off-by: Minchan Kim <minchan@xxxxxxxxxx>
> ---
> It's very early version which show the concept so I still marked it with RFC.
> I just tested it with simple test and works.
> This patch is needed indepth review from memory-hotplug guys from fujitsu
> because I saw there are lots of patches recenlty they sent to about
> memory-hotplug change. Please take a look at this patch.

[...]

> @@ -948,8 +954,13 @@ static int move_freepages(struct zone *zone,
>  		}
>  
>  		order = page_order(page);
> -		list_move(&page->lru,
> -			  &zone->free_area[order].free_list[migratetype]);
> +		if (migratetype != MIGRATE_ISOLATE) {
> +			list_move(&page->lru,
> +				&zone->free_area[order].free_list[migratetype]);
> +		} else {
> +			list_del(&page->lru);
> +			isolate_free_page(page, order);
> +		}
>  		page += 1 << order;
>  		pages_moved += 1 << order;
>  	}

Shouldn't NR_FREE_PAGES counter be decreased somewhere above?

[ I can see that it is not modified in __free_pages_ok() and
  free_hot_cold_page() because page is still counted as non-free one but
  here situation is different AFAICS. ]

I tested the patch locally here with CONFIG_CMA=y and it causes some
major problems for CMA (multiple errors from dma_alloc_from_contiguous()
about memory ranges being busy and allocation failures).

[ I'm sorry that I don't know more details yet but the issue should be
  easily reproducible. ]

Best regards,
--
Bartlomiej Zolnierkiewicz
Samsung Poland R&D Center
--
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/