[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <d692865c-82ae-103f-b48e-9b7682de28b6@163.com>
Date: Thu, 11 Feb 2021 19:01:02 +0800
From: zhou xianrong <xianrong_zhou@....com>
To: Michal Hocko <mhocko@...e.com>
Cc: iamjoonsoo.kim@....com, linux-mm@...ck.org,
linux-kernel@...r.kernel.org, akpm@...ux-foundation.org,
rostedt@...dmis.org, mingo@...hat.com, vbabka@...e.cz,
rientjes@...gle.com, willy@...ux.intel.com,
pankaj.gupta.linux@...il.com, bhe@...hat.com, ying.huang@...el.com,
minchan@...nel.org, ruxian.feng@...nssion.com,
kai.cheng@...nssion.com, zhao.xu@...nssion.com,
yunfeng.lan@...nssion.com, zhouxianrong@....com,
zhou xianrong <xianrong.zhou@...nssion.com>
Subject: Re: [PATCH] kswapd: no need reclaim cma pages triggered by unmovable
allocation
On 2021/2/10 下午9:14, Michal Hocko wrote:
> On Wed 10-02-21 12:07:57, zhou xianrong wrote:
>> On 2021/2/9 下午5:23, Michal Hocko wrote:
>>> On Tue 09-02-21 16:23:13, zhou wrote:
>>>> From: zhou xianrong <xianrong.zhou@...nssion.com>
>>>>
>>>> For purpose of better migration cma pages are allocated after
>>>> failure movalbe allocations and are used normally for file pages
>>>> or anonymous pages.
>>>>
>>>> In reclaim path so many cma pages if configurated are reclaimed
>>>> from lru lists in kswapd mainly or direct reclaim triggered by
>>>> unmovable or reclaimable allocations. But these cma pages can not
>>>> be used by original unmovable or reclaimable allocations. So the
>>>> reclaim are unnecessary.
>>>>
>>>> In a same system if the cma pages were configurated to large then
>>>> more failture unmovable (vmalloc etc.) or reclaimable (slab etc.)
>>>> allocations are arised and then more kswapd rounds are triggered
>>>> and then more cma pages are reclaimed.
>>> Could you be more specific? Do you have any numbers and an example
>>> configuration when this is visible?
>> It should be implicit.
> Right but the scale of the problem is an important part of _any_ patch
> justification.
Sorry. The relative description is not suitable and should be removed.
Powered by blists - more mailing lists