Re: [PATCH] mm: remove zone->prev_prioriy

[KOSAKI Motohiro]

> On Wed, 11 Feb 2009 20:06:46 +0900 (JST) KOSAKI Motohiro <kosaki.motohiro@xxxxxxxxxxxxxx> wrote:
> 
> > > On Tue, 10 Feb 2009 19:57:01 +0900
> > > MinChan Kim <minchan.kim@xxxxxxxxx> wrote:
> > > 
> > > > As you know, prev_priority is used as a measure of how much stress page reclaim.
> > > > But now we doesn't need it due to split-lru's way.
> > > > 
> > > > I think it would be better to remain why prev_priority isn't needed any more
> > > > and how split-lru can replace prev_priority's role in changelog.
> > > > 
> > > > In future, it help mm newbies understand change history, I think.
> > > 
> > > Yes, I'd be fascinated to see that explanation.
> > > 
> > > In http://groups.google.pn/group/linux.kernel/browse_thread/thread/fea9c9a0b43162a1
> > > it was asserted that we intend to use prev_priority again in the future.
> > > 
> > > We discussed this back in November:
> > > http://lkml.indiana.edu/hypermail/linux/kernel/0811.2/index.html#00001
> > > 
> > > And I think that I still think that the VM got worse due to its (new)
> > > failure to track previous state.  IIRC, the response to that concern
> > > was quite similar to handwavy waffling.
> > 
> > Yes.
> > I still think it's valuable code.
> > I think, In theory, VM sould take parallel reclaim bonus.
> 
> prev_priority had nothing to do with concurrent reclaim?
> 
> It was there so that when a task enters direct reclaim against a zone,
> it will immediately adopt the state which the task which most recently
> ran direct reclaim had.
> 
> Without this feature, each time a task enters direct reclaim it will need
> to "relearn" that state - ramping up, making probably-incorrect
> decisions as it does so.

Yes, I perfectly agree to you.
theorically, prev_priority is very valuable stuff.

rest only problem is, I should found good workload and re-integrate
prev_priority to reclaim code.

I (and many VM people) strongly dislike any regression.
then, if I can't find good workload, I can't change any VM behavior.

Do you have any suggestion?



--
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/