Re: [RFC Patch] mm/vmscan.c: not inherit classzone_idx from previous reclaim

[Mel Gorman]

On Wed, Feb 12, 2020 at 10:25:55AM +0800, Wei Yang wrote:
> On Tue, Feb 11, 2020 at 10:42:23AM +0000, Mel Gorman wrote:
> >On Sun, Feb 09, 2020 at 03:41:45PM +0800, Wei Yang wrote:
> >> Before commit e716f2eb24de ("mm, vmscan: prevent kswapd sleeping
> >> prematurely due to mismatched classzone_idx"), classzone_idx could have
> >> two possibilities on a new loop based on whether there is a wakeup
> >> during reclaiming:
> >> 
> >>   * 0 if no wakeup
> >>   * the classzone_idx request by wakeup
> >> 
> >> As described in the changelog, this commit is willing to change the
> >> first case to (MAX_NR_ZONES - 1) to avoid some premature sleep. But it
> >> does not achieve the goal.
> >> 
> >> There are two versions of kswapd_classzone_idx() since this change:
> >> 
> >>   * commit e716f2eb24de ("mm, vmscan: prevent kswapd sleeping
> >>     prematurely due to mismatched classzone_idx")
> >>   * commit dffcac2cb88e ("mm/vmscan.c: prevent useless kswapd loops")
> >> 
> >> Both of them would return the classzone_idx we passed as the 2nd
> >> parameter when (pgdat->kswapd_classzone_idx == MAX_NR_ZONES). This
> >> means if there is no wakeup during reclaiming, we would use
> >> classzone_idx in previous round to sleep.
> >> 
> >
> >This is somewhat intended.
> >
> >> This patch fixes the logic by using (MAX_NR_ZONES - 1) for the first
> >> case.
> >> 
> >
> >Ok, what is the user-visible impact that is fixed by this patch or is
> >this based on code review only? Please describe the test case exactly
> >and the before and after results. I ask because this area is a magnet for
> >regressions and intuitive ideas often lead to counter-intuitive results.
> >
> 
> This is based on code review only. I know your concern. This is an area more
> like art then engineering :-)
> 

Then I'm afraid that until there is a corner case identified and a
description of the impact it's

Nacked-by: Mel Gorman <mgorman@xxxxxxxxxxxxxxxxxxx>

> Would you mind sharing some idea why we intend to inherit the classzone_idx?
> And for kswapd_order, we would restart at 0 if no wakeup during reclaim.
> 

Broadly speaking it was driven by cases whereby kswapd either a) fell
asleep prematurely and there were many stalls in direct reclaim before
kswapd recovered, b) stalls in direct reclaim immediately after kswapd went
to sleep or c) kswapd reclaimed for lower zones and went to sleep while
parallel tasks were direct reclaiming in higher zones or higher orders.

-- 
Mel Gorman
SUSE Labs