[dm-devel] [PATCH 15/15] dm-zoned: prefer full zones for reclaim

[Damien Le Moal]

On Wed, 2020-05-27 at 08:22 +0200, Hannes Reinecke wrote:
> Prefer full zones when selecting the next zone for reclaim.
> 
> Signed-off-by: Hannes Reinecke <hare at suse.de>
> ---
>  drivers/md/dm-zoned-metadata.c | 9 ++++++++-
>  1 file changed, 8 insertions(+), 1 deletion(-)
> 
> diff --git a/drivers/md/dm-zoned-metadata.c b/drivers/md/dm-zoned-metadata.c
> index b89b3d3b9ec9..f161ef4e3d3d 100644
> --- a/drivers/md/dm-zoned-metadata.c
> +++ b/drivers/md/dm-zoned-metadata.c
> @@ -1936,7 +1936,7 @@ static struct dm_zone *dmz_get_rnd_zone_for_reclaim(struct dmz_metadata *zmd,
>  						    unsigned int idx, bool idle)
>  {
>  	struct dm_zone *dzone = NULL;
> -	struct dm_zone *zone;
> +	struct dm_zone *zone, *last = NULL;
>  	struct list_head *zone_list;
>  
>  	/* If we have cache zones select from the cache zone list */
> @@ -1953,6 +1953,13 @@ static struct dm_zone *dmz_get_rnd_zone_for_reclaim(struct dmz_metadata *zmd,
>  			dzone = zone->bzone;
>  			if (dzone->dev->dev_idx != idx)
>  				continue;
> +			if (!last) {
> +				last = dzone;
> +				continue;
> +			}
> +			if (last->weight < dzone->weight)
> +				continue;
> +			dzone = last;
>  		} else
>  			dzone = zone;
>  		if (dmz_lock_zone_reclaim(dzone))

If all random/cache zones are used but none of them satisfy the
condition last->weight < dzone->weight, we may end up starving reclaim
and having user IOs accessing a new chunk wait a loooong time, if not
forever, No ? I agree that aiming at reclaim of full zones first is
more efficient, but we need a fallback to ensure forward progress.

-- 
Damien Le Moal
Western Digital Research