Re: [PATCH 1/1] mm: vmscan: Reduce throttling due to a failure to make progress

[Mel Gorman]

On Wed, Dec 01, 2021 at 02:52:01PM +0100, Mike Galbraith wrote:
> On Wed, 2021-12-01 at 13:01 +0000, Mel Gorman wrote:
> > On Tue, Nov 30, 2021 at 06:59:58PM +0100, Mike Galbraith wrote:
> > > On Tue, 2021-11-30 at 17:27 +0000, Mel Gorman wrote:
> > > >
> > > > Obviously a fairly different experience and most likely due to
> > > > the
> > > > underlying storage.
> > >
> > > I bet a virtual nickle this is the sore spot.
> > >
> >
> > You win a virtual nickle!
> 
> I'm rich I'm rich... oh dang, virtual.
> 
> I went back to 5.15, and confirmed that wait_iff_congested() did not
> ever sleep with the try to eat /dev/zero load.  Nor did it with insane
> overcommit swap storm from hell with as much IO going on as my little
> box is capable of generating, making the surrounding congestion bits
> look.. down right expendable.
> 

wait_iff_congested was broken once the block layer stopped tracking
congestion and became a glorified cond_resched() in most cases. This is
why the series aimed to remove the reliance on
congestion_wait/wait_iff_congested.

-- 
Mel Gorman
SUSE Labs