rbd: set io_min, io_opt and discard_granularity to alloc_size

idryomov · idryomov · commit 16d80c54ad42 · 2019-03-18T20:30:01.000+01:00
Now that we have alloc_size that controls our discard behavior, it
doesn't make sense to have these set to object (set) size.  alloc_size
defaults to 64k, but because discard_granularity is likely 4M, only
ranges that are equal to or bigger than 4M can be considered during
fstrim.  A smaller io_min is also more likely to be met, resulting in
fewer deferred writes on bluestore OSDs.

Signed-off-by: Ilya Dryomov &lt;idryomov@gmail.com&gt;
Reviewed-by: Jason Dillaman &lt;dillaman@redhat.com&gt;
diff --git a/drivers/block/rbd.c b/drivers/block/rbd.c
@@ -833,7 +833,7 @@ static int parse_rbd_opts_token(char *c, void *private)
 		pctx->opts->queue_depth = intval;
 		break;
 	case Opt_alloc_size:
-		if (intval < 1) {
+		if (intval < SECTOR_SIZE) {
 			pr_err("alloc_size out of range\n");
 			return -EINVAL;
 		}
@@ -4203,12 +4203,12 @@ static int rbd_init_disk(struct rbd_device *rbd_dev)
 	q->limits.max_sectors = queue_max_hw_sectors(q);
 	blk_queue_max_segments(q, USHRT_MAX);
 	blk_queue_max_segment_size(q, UINT_MAX);
-	blk_queue_io_min(q, objset_bytes);
-	blk_queue_io_opt(q, objset_bytes);
+	blk_queue_io_min(q, rbd_dev->opts->alloc_size);
+	blk_queue_io_opt(q, rbd_dev->opts->alloc_size);
 
 	if (rbd_dev->opts->trim) {
 		blk_queue_flag_set(QUEUE_FLAG_DISCARD, q);
-		q->limits.discard_granularity = objset_bytes;
+		q->limits.discard_granularity = rbd_dev->opts->alloc_size;
 		blk_queue_max_discard_sectors(q, objset_bytes >> SECTOR_SHIFT);
 		blk_queue_max_write_zeroes_sectors(q, objset_bytes >> SECTOR_SHIFT);
 	}
