Re: 103 sec. latency: sync_page() with TASK_UNINTERRUPTIBLE (?) (bisected)

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

Hi

> > I bisected it down to f1a9ee758de7de1e040de849fdef46e6802ea117
> > ("kswapd should only wait on IO if there is IO").
> > 
> > If a revert the above commit from the latest git kernel
> > (v2.6.25-rc6-243-g028011e), then I get an very good interactivity
> > while xjed is loading the 380MB file (meaning that I can do other things
> > during the load). Otherwise it takes almost 3 times longer to load,
> > and interactivity is very bad in the meanwhile. I doubled checked it.
> > 
> > I hope this information will be useful, and I can try any patches
> > if necessary.
> > 
> 
> Thanks, I queued the below reversion.  It would be great if you could
> confirm that this patch does indeed make 2.6.25-rc6 work as well as 2.6.24.
> 
> From: Andrew Morton <akpm@linux-foundation.org>
> 
> Revert
> 
> commit f1a9ee758de7de1e040de849fdef46e6802ea117
> Author: Rik van Riel <riel@redhat.com>
> Date:   Thu Feb 7 00:14:08 2008 -0800
> 
>     kswapd should only wait on IO if there is IO

if you don't mind, I hope join to test to this patch.
I think his basic idea is good.
but unfortunately, it has corner case problem.
if do more test and remove any regression, this patch is valuable
for merge again.


- kosaki