Re: [PATCH] mm/vmscan: don't scan adjust too much if current is not kswapd

[Andrew Morton]

On Wed, 14 Sep 2022 10:33:18 +0800 Hongchen Zhang <zhanghongchen@xxxxxxxxxxx> wrote:

> when a process falls into page fault and there is not enough free
> memory,it will do direct reclaim. At the same time,it is holding
> mmap_lock.So in case of multi-thread,it should exit from page fault
> ASAP.
> When reclaim memory,we do scan adjust between anon and file lru which
> may cost too much time and trigger hung task for other thread.So for a
> process which is not kswapd,it should just do a little scan adjust.

Well, that's a pretty nasty bug.  Before diving into a possible fix,
can you please tell us more about how this happens?  What sort of
machine, what sort of workload.  Can you suggest why others are not
experiencing this?

> --- a/mm/vmscan.c
> +++ b/mm/vmscan.c
> @@ -3042,11 +3042,17 @@ static void shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
>  		nr[lru] = targets[lru] * (100 - percentage) / 100;
>  		nr[lru] -= min(nr[lru], nr_scanned);
>  
> +		if (!current_is_kswapd())
> +			nr[lru] = min(nr[lru], nr_to_reclaim);
> +
>  		lru += LRU_ACTIVE;
>  		nr_scanned = targets[lru] - nr[lru];
>  		nr[lru] = targets[lru] * (100 - percentage) / 100;
>  		nr[lru] -= min(nr[lru], nr_scanned);
>  
> +		if (!current_is_kswapd())
> +			nr[lru] = min(nr[lru], nr_to_reclaim);
> +
>  		scan_adjusted = true;
>  	}
>  	blk_finish_plug(&plug);

It would be better if these additions had code comments explaining why
they're there.  But let's more fully understand the problem before
altering your patch.