Re: [Patch] Call cond_resched() at bottom of main look in balance_pgdat()

[Minchan Kim]

On Wed, Jun 23, 2010 at 6:33 AM, Johannes Weiner <hannes@xxxxxxxxxxx> wrote:
> On Tue, Jun 22, 2010 at 01:29:17PM +0900, Minchan Kim wrote:
>> On Tue, Jun 22, 2010 at 12:23 PM, KOSAKI Motohiro
>> <kosaki.motohiro@xxxxxxxxxxxxxx> wrote:
>> >> >> Kosaki's patch's goal is that kswap doesn't yield cpu if the zone doesn't meet its
>> >> >> min watermark to avoid failing atomic allocation.
>> >> >> But this patch could yield kswapd's time slice at any time.
>> >> >> Doesn't the patch break your goal in bb3ab59683?
>> >> >
>> >> > No. it don't break.
>> >> >
>> >> > Typically, kswapd periodically call shrink_page_list() and it call
>> >> > cond_resched() even if bb3ab59683 case.
>> >>
>> >> Hmm. If it is, bb3ab59683 is effective really?
>> >>
>> >> The bb3ab59683's goal is prevent CPU yield in case of free < min_watermark.
>> >> But shrink_page_list can yield cpu from kswapd at any time.
>> >> So I am not sure what is bb3ab59683's benefit.
>> >> Did you have any number about bb3ab59683's effectiveness?
>> >> (Of course, I know it's very hard. Just out of curiosity)
>> >>
>> >> As a matter of fact, when I saw this Larry's patch, I thought it would
>> >> be better to revert bb3ab59683. Then congestion_wait could yield CPU
>> >> to other process.
>> >>
>> >> What do you think about?
>> >
>> > No. The goal is not prevent CPU yield. The goal is avoid unnecessary
>> > _long_ sleep (i.e. congestion_wait(BLK_RW_ASYNC, HZ/10)).
>>
>> I meant it.
>>
>> > Anyway we can't refuse CPU yield on UP. it lead to hangup ;)
>> >
>> > What do you mean the number? If it mean how much reduce congestion_wait(),
>> > it was posted a lot of time. If it mean how much reduce page allocation
>> > failure bug report, I think it has been observable reduced since half
>> > years ago.
>>
>> I meant second.
>> Hmm. I doubt it's observable since at that time, Mel had posted many
>> patches to reduce page allocation fail. bb3ab59683 was just one of
>> them.
>>
>> >
>> > If you have specific worried concern, can you please share it?
>> >
>>
>> My concern is that I don't want to add new band-aid on uncertain
>> feature to solve
>> regression of uncertain feature.(Sorry for calling Larry's patch as band-aid.).
>> If we revert bb3ab59683, congestion_wait in balance_pgdat could yield
>> cpu from kswapd.
>>
>> If you insist on bb3ab59683's effective and have proved it at past, I
>> am not against it.
>>
>> And If it's regression of bb3ab59683, Doesn't it make sense following as?
>> It could restore old behavior.
>>
>> ---
>> Â Â Â Â Â Â Â Â Â* OK, kswapd is getting into trouble. ÂTake a nap, then take
>> Â Â Â Â Â Â Â Â Â* another pass across the zones.
>> Â Â Â Â Â Â Â Â Â*/
>> Â Â Â Â Â Â Â Â if (total_scanned && (priority < DEF_PRIORITY - 2)) {
>> Â Â Â Â Â Â Â Â Â Â Â Â if (has_under_min_watermark_zone) {
>> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â count_vm_event(KSWAPD_SKIP_CONGESTION_WAIT);
>> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â /* allowing CPU yield to go on
>> watchdog or OOMed task */
>> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â cond_resched();
>
> We have two things here: one is waiting for some IO to complete, which
> we skip if we are in a hurry. ÂThe other thing is that we have a
> potentially long-running loop with no garuanteed rescheduling point in
> it. ÂI would rather not mix up those two and let this cond_resched()
> for #2 stand on it's own and be self-explanatory.
>
> So,
>
> Acked-by: Johannes Weiner <hannes@xxxxxxxxxxx>
>
> to Larry's patch (or KOSAKI-san's version of it for that matter).
>

Okay. As I hear Kosaki and Hannes opinions, I was paranoid.
Thanks for good comment!, Kosaki and Hannes.
Feel free to add my sign to Kosaki's version(I like detailed description :) )

Reviewed-by: Minchan Kim <minchan.kim@xxxxxxxxx>


-- 
Kind 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/