Disable LBA weighting on files and SSDs

The LBA weighting makes sense on rotational media where the outer tracks
have twice the bandwidth of the inner tracks. However, it is detrimental
on nonrotational media such as solid state disks, where the only effect
is to ensure that metaslabs enter the best-fit allocation behavior
sooner, which is detrimental to performance. It also makes no sense on
files where the underlying filesystem can arrange things however it
wants.

Signed-off-by: Richard Yao <[email protected]>
Signed-off-by: Brian Behlendorf <[email protected]>
Closes #3712

5 files changed, 17 insertions, 2 deletions

diff --git a/include/sys/vdev_impl.h b/include/sys/vdev_impl.h
index 6b27e75ae..1371a3f03 100644
--- a/include/sys/vdev_impl.h
+++ b/include/sys/vdev_impl.h
@@ -151,6 +151,7 @@ struct vdev {
 	vdev_stat_t	vdev_stat;	/* virtual device statistics	*/
 	boolean_t	vdev_expanding;	/* expand the vdev?		*/
 	boolean_t	vdev_reopening;	/* reopen in progress?		*/
+	boolean_t	vdev_nonrot;	/* true if solid state		*/
 	int		vdev_open_error; /* error on last open		*/
 	kthread_t	*vdev_open_thread; /* thread opening children	*/
 	uint64_t	vdev_crtxg;	/* txg when top-level was added */
diff --git a/module/zfs/metaslab.c b/module/zfs/metaslab.c
index b328cbb0a..59bcefd34 100644
--- a/module/zfs/metaslab.c
+++ b/module/zfs/metaslab.c
@@ -1518,7 +1518,7 @@ metaslab_weight(metaslab_t *msp)
 	 * In effect, this means that we'll select the metaslab with the most
 	 * free bandwidth rather than simply the one with the most free space.
 	 */
-	if (metaslab_lba_weighting_enabled) {
+	if (!vd->vdev_nonrot && metaslab_lba_weighting_enabled) {
 		weight = 2 * weight - (msp->ms_id * weight) / vd->vdev_ms_count;
 		ASSERT(weight >= space && weight <= 2 * space);
 	}
diff --git a/module/zfs/vdev.c b/module/zfs/vdev.c
index 8e50ababc..7aff5455b 100644
--- a/module/zfs/vdev.c
+++ b/module/zfs/vdev.c
@@ -1108,6 +1108,7 @@ vdev_open_child(void *arg)
 	vd->vdev_open_thread = curthread;
 	vd->vdev_open_error = vdev_open(vd);
 	vd->vdev_open_thread = NULL;
+	vd->vdev_parent->vdev_nonrot &= vd->vdev_nonrot;
 }
 
 static boolean_t
@@ -1134,15 +1135,19 @@ vdev_open_children(vdev_t *vd)
 	int children = vd->vdev_children;
 	int c;
 
+	vd->vdev_nonrot = B_TRUE;
+
 	/*
 	 * in order to handle pools on top of zvols, do the opens
 	 * in a single thread so that the same thread holds the
 	 * spa_namespace_lock
 	 */
 	if (vdev_uses_zvols(vd)) {
-		for (c = 0; c < children; c++)
+		for (c = 0; c < children; c++) {
 			vd->vdev_child[c]->vdev_open_error =
 			    vdev_open(vd->vdev_child[c]);
+			vd->vdev_nonrot &= vd->vdev_child[c]->vdev_nonrot;
+		}
 		return;
 	}
 	tq = taskq_create("vdev_open", children, minclsyspri,
@@ -1153,6 +1158,9 @@ vdev_open_children(vdev_t *vd)
 		    TQ_SLEEP) != 0);
 
 	taskq_destroy(tq);
+
+	for (c = 0; c < children; c++)
+		vd->vdev_nonrot &= vd->vdev_child[c]->vdev_nonrot;
 }
 
 /*
diff --git a/module/zfs/vdev_disk.c b/module/zfs/vdev_disk.c
index eb77c269c..380ede35b 100644
--- a/module/zfs/vdev_disk.c
+++ b/module/zfs/vdev_disk.c
@@ -301,6 +301,9 @@ skip_open:
 	/* Clear the nowritecache bit, causes vdev_reopen() to try again. */
 	v->vdev_nowritecache = B_FALSE;
 
+	/* Inform the ZIO pipeline that we are non-rotational */
+	v->vdev_nonrot = blk_queue_nonrot(bdev_get_queue(vd->vd_bdev));
+
 	/* Physical volume size in bytes */
 	*psize = bdev_capacity(vd->vd_bdev);
 
diff --git a/module/zfs/vdev_file.c b/module/zfs/vdev_file.c
index e61240fdc..a29ea7bf9 100644
--- a/module/zfs/vdev_file.c
+++ b/module/zfs/vdev_file.c
@@ -57,6 +57,9 @@ vdev_file_open(vdev_t *vd, uint64_t *psize, uint64_t *max_psize,
 	vattr_t vattr;
 	int error;
 
+	/* Rotational optimizations only make sense on block devices */
+	vd->vdev_nonrot = B_TRUE;
+
 	/*
 	 * We must have a pathname, and it must be absolute.
 	 */