vdev_mirror: load balancing fixes

vdev_queue:
- Track the last position of each vdev, including the io size,
  in order to detect linear access of the following zio.
- Remove duplicate `vq_lastoffset`

vdev_mirror:
- Correctly calculate the zio offset (signedness issue)
- Deprecate `vdev_queue_register_lastoffset()`
- Add `VDEV_LABEL_START_SIZE` to zio offset of leaf vdevs

Reviewed-by: Brian Behlendorf <[email protected]>
Signed-off-by: Gvozden Neskovic <[email protected]>
Closes #6461

2 files changed, 23 insertions, 34 deletions

diff --git a/module/zfs/vdev_mirror.c b/module/zfs/vdev_mirror.c
index 15d1f204f..0cf2bc665 100644
--- a/module/zfs/vdev_mirror.c
+++ b/module/zfs/vdev_mirror.c
@@ -116,7 +116,8 @@ static const zio_vsd_ops_t vdev_mirror_vsd_ops = {
 static int
 vdev_mirror_load(mirror_map_t *mm, vdev_t *vd, uint64_t zio_offset)
 {
-	uint64_t lastoffset;
+	uint64_t last_offset;
+	int64_t offset_diff;
 	int load;
 
 	/* All DVAs have equal weight at the root. */
@@ -129,13 +130,17 @@ vdev_mirror_load(mirror_map_t *mm, vdev_t *vd, uint64_t zio_offset)
 	 * worse overall when resilvering with compared to without.
 	 */
 
+	/* Fix zio_offset for leaf vdevs */
+	if (vd->vdev_ops->vdev_op_leaf)
+		zio_offset += VDEV_LABEL_START_SIZE;
+
 	/* Standard load based on pending queue length. */
 	load = vdev_queue_length(vd);
-	lastoffset = vdev_queue_lastoffset(vd);
+	last_offset = vdev_queue_last_offset(vd);
 
 	if (vd->vdev_nonrot) {
 		/* Non-rotating media. */
-		if (lastoffset == zio_offset)
+		if (last_offset == zio_offset)
 			return (load + zfs_vdev_mirror_non_rotating_inc);
 
 		/*
@@ -148,16 +153,16 @@ vdev_mirror_load(mirror_map_t *mm, vdev_t *vd, uint64_t zio_offset)
 	}
 
 	/* Rotating media I/O's which directly follow the last I/O. */
-	if (lastoffset == zio_offset)
+	if (last_offset == zio_offset)
 		return (load + zfs_vdev_mirror_rotating_inc);
 
 	/*
 	 * Apply half the seek increment to I/O's within seek offset
-	 * of the last I/O queued to this vdev as they should incur less
+	 * of the last I/O issued to this vdev as they should incur less
 	 * of a seek increment.
 	 */
-	if (ABS(lastoffset - zio_offset) <
-	    zfs_vdev_mirror_rotating_seek_offset)
+	offset_diff = (int64_t)(last_offset - zio_offset);
+	if (ABS(offset_diff) < zfs_vdev_mirror_rotating_seek_offset)
 		return (load + (zfs_vdev_mirror_rotating_seek_inc / 2));
 
 	/* Apply the full seek increment to all other I/O's. */
@@ -382,29 +387,20 @@ vdev_mirror_child_select(zio_t *zio)
 		mm->mm_preferred_cnt++;
 	}
 
-	if (mm->mm_preferred_cnt == 1) {
-		vdev_queue_register_lastoffset(
-		    mm->mm_child[mm->mm_preferred[0]].mc_vd, zio);
+	if (mm->mm_preferred_cnt == 1)
 		return (mm->mm_preferred[0]);
-	}
 
-	if (mm->mm_preferred_cnt > 1) {
-		int c = vdev_mirror_preferred_child_randomize(zio);
 
-		vdev_queue_register_lastoffset(mm->mm_child[c].mc_vd, zio);
-		return (c);
-	}
+	if (mm->mm_preferred_cnt > 1)
+		return (vdev_mirror_preferred_child_randomize(zio));
 
 	/*
 	 * Every device is either missing or has this txg in its DTL.
 	 * Look for any child we haven't already tried before giving up.
 	 */
 	for (c = 0; c < mm->mm_children; c++) {
-		if (!mm->mm_child[c].mc_tried) {
-			vdev_queue_register_lastoffset(mm->mm_child[c].mc_vd,
-			    zio);
+		if (!mm->mm_child[c].mc_tried)
 			return (c);
-		}
 	}
 
 	/*
diff --git a/module/zfs/vdev_queue.c b/module/zfs/vdev_queue.c
index 6b3e87291..40cba340a 100644
--- a/module/zfs/vdev_queue.c
+++ b/module/zfs/vdev_queue.c
@@ -393,7 +393,7 @@ vdev_queue_init(vdev_t *vd)
 		    sizeof (zio_t), offsetof(struct zio, io_queue_node));
 	}
 
-	vq->vq_lastoffset = 0;
+	vq->vq_last_offset = 0;
 }
 
 void
@@ -699,9 +699,8 @@ again:
 	 */
 	tree = vdev_queue_class_tree(vq, p);
 	vq->vq_io_search.io_timestamp = 0;
-	vq->vq_io_search.io_offset = vq->vq_last_offset + 1;
-	VERIFY3P(avl_find(tree, &vq->vq_io_search,
-	    &idx), ==, NULL);
+	vq->vq_io_search.io_offset = vq->vq_last_offset - 1;
+	VERIFY3P(avl_find(tree, &vq->vq_io_search, &idx), ==, NULL);
 	zio = avl_nearest(tree, idx, AVL_AFTER);
 	if (zio == NULL)
 		zio = avl_first(tree);
@@ -728,7 +727,7 @@ again:
 	}
 
 	vdev_queue_pending_add(vq, zio);
-	vq->vq_last_offset = zio->io_offset;
+	vq->vq_last_offset = zio->io_offset + zio->io_size;
 
 	return (zio);
 }
@@ -806,7 +805,7 @@ vdev_queue_io_done(zio_t *zio)
 }
 
 /*
- * As these three methods are only used for load calculations we're not
+ * As these two methods are only used for load calculations we're not
  * concerned if we get an incorrect value on 32bit platforms due to lack of
  * vq_lock mutex use here, instead we prefer to keep it lock free for
  * performance.
@@ -818,15 +817,9 @@ vdev_queue_length(vdev_t *vd)
 }
 
 uint64_t
-vdev_queue_lastoffset(vdev_t *vd)
+vdev_queue_last_offset(vdev_t *vd)
 {
-	return (vd->vdev_queue.vq_lastoffset);
-}
-
-void
-vdev_queue_register_lastoffset(vdev_t *vd, zio_t *zio)
-{
-	vd->vdev_queue.vq_lastoffset = zio->io_offset + zio->io_size;
+	return (vd->vdev_queue.vq_last_offset);
 }
 
 #if defined(_KERNEL) && defined(HAVE_SPL)