Re: [PATCH] prevent to reclaim anon page of lumpy reclaim for no swap space

[Minchan Kim]

Hi, Lee.

On Thu, Jun 25, 2009 at 11:54 PM, Lee
Schermerhorn<Lee.Schermerhorn@xxxxxx> wrote:
> On Thu, 2009-06-25 at 23:44 +0900, Minchan Kim wrote:
>> On Thu, Jun 25, 2009 at 11:14 PM, KOSAKI
>> Motohiro<kosaki.motohiro@xxxxxxxxxxxxxx> wrote:
>> >> This patch prevent to reclaim anon page in case of no swap space.
>> >> VM already prevent to reclaim anon page in various place.
>> >> But it doesnt't prevent it for lumpy reclaim.
>> >>
>> >> It shuffles lru list unnecessary so that it is pointless.
>> >
>> > NAK.
>> >
>> > 1. if system have no swap, add_to_swap() never get swap entry.
>> > Â eary check don't improve performance so much.
>>
>> Hmm. I mean no swap space but not no swap device.
>> add_to_swap ? You mean Rik pointed me out ?
>> If system have swap device, Rik's pointing is right.
>> I will update his suggestion.
>>
>> > 2. __isolate_lru_page() is not only called lumpy reclaim case, but
>> > also be called
>> > Â Ânormal reclaim.
>>
>> You mean about performance degradation ?
>> I think most case have enough swap space and then one condition
>> variable(nr_swap_page) check is trivial. I think.
>> We can also use [un]likely but I am not sure it help us.
>>
>>
>> > 3. if system have no swap, anon pages shuffuling doesn't cause any matter.
>>
>> Again, I mean no swap space but no swap device system.
>> And I have a plan to remove anon_vma in no swap device system.
>>
>> As you point me out, it's pointless in no swap device system.
>> I don't like unnecessary structure memory footprint and locking overhead.
>> I think no swap device system is problem in server environment as well
>> as embedded. but I am not sure when I will do. :)
>>
>
> How will we walk the reverse map for try_to_unmap() for page migration
> or try_to_munlock() w/o anon_vma? ÂPerhaps one can remove anon_vma when
> there is no swap device and migration and the unevictable lru are not
> configured--e.g., for embedded systems.

You're right. In addition, there are HWPoison and maybe KSM.
Also, unevictable lru list isn't option any more even embedded system.

Actually I considered it in embedded system as you said.
I think above enumerated cases are not needed in embedded system.
Memory footprint and unnecessary locking is more important in embedded
since small memory and realtime.

Anyway, I think it's not easy so now it's just plan.
Welcome to any comment to prevent my vain effort. :)

Thanks for valuable comment. Lee. :)

> Lee
>
>



-- 
Kinds regards,
Minchan Kim
--
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/