Re: [PATCH] mm: stop kswapd's infinite loop at high order allocation take2

["KOSAKI Motohiro" <kosaki.motohiro@jp.fujitsu.com>]

> Hi, kosaki-san.
>
> I read the previous threads now. It's rather late :(.
>
> I think it's rather awkward that sudden big change of order from 10 to 0.
>
> This problem causes zone_water_mark's fail.
> It mean now this zone's proportional free page per order size is not good.
> Although order-0 page is very important, Shouldn't we consider other
> order allocations ?
>
> So I want to balance zone's proportional free page.
> How about following ?
>
> if (nr_reclaimed < SWAP_CLUSTER_MAX) {
>   if (order != 0) {
>     order -=1;
>     sc.order -=1;
>   }
> }
>
> It prevents infinite loop and do best effort to make zone's
> proportional free page per order size good.
>
> It's just my opinion within my knowledge.
> If it have a problem, pz, explain me :)

Please read Nick's expalin. it explain very kindly :)