Re: [PATCH 1/5] mm/vmscan: Throttle reclaim until some writeback completes if congested

[Mel Gorman]

On Wed, Sep 22, 2021 at 08:16:20PM +0800, Hillf Danton wrote:
> On Mon, 20 Sep 2021 09:54:32 +0100 Mel Gorman wrote:
> > +static void
> > +reclaim_throttle(pg_data_t *pgdat, enum vmscan_throttle_state reason,
> > +							long timeout)
> > +{
> > +	wait_queue_head_t *wqh =3D &pgdat->reclaim_wait;
> > +	unsigned long start =3D jiffies;
> > +	long ret;
> > +	DEFINE_WAIT(wait);
> > +
> > +	atomic_inc(&pgdat->nr_reclaim_throttled);
> > +	WRITE_ONCE(pgdat->nr_reclaim_start,
> > +		 node_page_state(pgdat, NR_THROTTLED_WRITTEN));
> 
> Missing wakeup could happen if the current sleeper overwrites 
> pgdat->nr_reclaim_start set by the existing sleeper.
> 
> 	if (1 == atomic_inc_and_return(&pgdat->nr_reclaim_throttled))
> 		WRITE_ONCE(pgdat->nr_reclaim_start,
> 				node_page_state(pgdat, NR_THROTTLED_WRITTEN));
> 

Good spot, will fix.

-- 
Mel Gorman
SUSE Labs