dm zoned: Fix zone reclaim trigger

Only triggering reclaim based on the percentage of unmapped cache
zones can fail to detect cases where reclaim is needed, e.g. if the
target has only 2 or 3 cache zones and only one unmapped cache zone,
the percentage of free cache zones is higher than
DMZ_RECLAIM_LOW_UNMAP_ZONES (30%) and reclaim does not trigger.

This problem, combined with the fact that dmz_schedule_reclaim() is
called from dmz_handle_bio() without the map lock held, leads to a
race between zone allocation and dmz_should_reclaim() result.
Depending on the workload applied, this race can lead to the write
path waiting forever for a free zone without reclaim being triggered.

Fix this by moving dmz_schedule_reclaim() inside dmz_alloc_zone()
under the map lock. This results in checking the need for zone reclaim
whenever a new data or buffer zone needs to be allocated.

Also fix dmz_reclaim_percentage() to always return 0 if the number of
unmapped cache (or random) zones is less than or equal to 1.

Suggested-by: Shin'ichiro Kawasaki <shinichiro.kawasaki@wdc.com>
Signed-off-by: Damien Le Moal <damien.lemoal@wdc.com>
Reviewed-by: Hannes Reinecke <hare@suse.de>
Signed-off-by: Mike Snitzer <snitzer@redhat.com>

1 files changed, 8 insertions, 1 deletions

diff --git a/drivers/md/dm-zoned-metadata.c b/drivers/md/dm-zoned-metadata.c
index 5cf6f5f552e0..b298fefb022e 100644
--- a/drivers/md/dm-zoned-metadata.c
+++ b/drivers/md/dm-zoned-metadata.c
@@ -2217,8 +2217,15 @@ struct dm_zone *dmz_alloc_zone(struct dmz_metadata *zmd, unsigned int dev_idx,
 {
 	struct list_head *list;
 	struct dm_zone *zone;
-	int i = 0;
+	int i;
+
+	/* Schedule reclaim to ensure free zones are available */
+	if (!(flags & DMZ_ALLOC_RECLAIM)) {
+		for (i = 0; i < zmd->nr_devs; i++)
+			dmz_schedule_reclaim(zmd->dev[i].reclaim);
+	}
 
+	i = 0;
 again:
 	if (flags & DMZ_ALLOC_CACHE)
 		list = &zmd->unmap_cache_list;