Add vdev property to disable vdev scheduler

Added vdev property to disable the vdev scheduler.
The intention behind this property is to improve IOPS
performance when using o_direct.

Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Alexander Motin <alexander.motin@TrueNAS.com>
Signed-off-by: MigeljanImeri <ImeriMigel@gmail.com>
Closes #17358

include/sys/fs/zfs.h

@@ -389,9 +389,22 @@ typedef enum {
 	VDEV_PROP_SIT_OUT,
 	VDEV_PROP_AUTOSIT,
 	VDEV_PROP_SLOW_IO_EVENTS,
+	VDEV_PROP_SCHEDULER,
 	VDEV_NUM_PROPS
 } vdev_prop_t;
 
+/*
+ * Different scheduling behaviors for vdev scheduler property.
+ * VDEV_SCHEDULER_AUTO = Let ZFS decide - currently use scheduler on HDDs only.
+ * VDEV_SCHEDULER_ON = Always queue.
+ * VDEV_SCHEDULER_OFF = Never queue.
+ */
+typedef enum {
+	VDEV_SCHEDULER_AUTO,
+	VDEV_SCHEDULER_ON,
+	VDEV_SCHEDULER_OFF
+} vdev_scheduler_type_t;
+
 /*
  * Dataset property functions shared between libzfs and kernel.
  */

include/sys/vdev_impl.h

@@ -425,6 +425,7 @@ struct vdev {
 	boolean_t	vdev_resilver_deferred;  /* resilver deferred */
 	boolean_t	vdev_kobj_flag; /* kobj event record */
 	boolean_t	vdev_attaching; /* vdev attach ashift handling */
+	boolean_t	vdev_is_blkdev; /* vdev is backed by block device */
 	vdev_queue_t	vdev_queue;	/* I/O deadline schedule queue	*/
 	spa_aux_vdev_t	*vdev_aux;	/* for l2cache and spares vdevs	*/
 	zio_t		*vdev_probe_zio; /* root of current probe	*/
@@ -473,6 +474,7 @@ struct vdev {
 	boolean_t	vdev_slow_io_events;
 	uint64_t	vdev_slow_io_n;
 	uint64_t	vdev_slow_io_t;
+	uint64_t	vdev_scheduler; /* control how I/Os are submitted */
 };
 
 #define	VDEV_PAD_SIZE		(8 << 10)

lib/libzfs/libzfs.abi

@@ -6255,7 +6255,8 @@
       <enumerator name='VDEV_PROP_SIT_OUT' value='52'/>
       <enumerator name='VDEV_PROP_AUTOSIT' value='53'/>
       <enumerator name='VDEV_PROP_SLOW_IO_EVENTS' value='54'/>
-      <enumerator name='VDEV_NUM_PROPS' value='55'/>
+      <enumerator name='VDEV_PROP_SCHEDULER' value='55'/>
+      <enumerator name='VDEV_NUM_PROPS' value='56'/>
     </enum-decl>
     <typedef-decl name='vdev_prop_t' type-id='1573bec8' id='5aa5c90c'/>
     <class-decl name='zpool_load_policy' size-in-bits='256' is-struct='yes' visibility='default' id='2f65b36f'>

man/man7/vdevprops.7

@@ -194,6 +194,23 @@ If this device should perform new allocations, used to disable a device
 when it is scheduled for later removal.
 See
 .Xr zpool-remove 8 .
+.It Sy scheduler Ns = Ns Sy auto Ns | Ns Sy on Ns | Ns Sy off
+Controls how I/O requests are added to the vdev queue when reading or
+writing to this vdev.
+This property can be set on leaf vdevs.
+The value of these properties do not persist across vdev replacement.
+.Bl -tag -compact -width "auto"
+.It Ar auto
+Let ZFS choose which scheduler it thinks will be best.
+Currently, the scheduler will queue I/O if the vdev is backed by a rotational
+block device or file, and not queue otherwise.
+.It Ar on
+Always adds I/O requests to the vdev queue.
+.It Ar off
+Never adds I/O requests to the vdev queue.
+This is not recommended for vdevs backed by spinning disks as it could
+result in starvation.
+.El
 .El
 .Ss User Properties
 In addition to the standard native properties, ZFS supports arbitrary user

module/os/freebsd/zfs/vdev_geom.c

@@ -968,6 +968,9 @@ skip_open:
 	else
 		vd->vdev_nonrot = B_FALSE;
 
+	/* Is backed by a block device. */
+	vd->vdev_is_blkdev = B_TRUE;
+
 	/* Set when device reports it supports TRIM. */
 	error = g_getattr("GEOM::candelete", cp, &has_trim);
 	vd->vdev_has_trim = (error == 0 && has_trim);

module/os/linux/zfs/vdev_disk.c

@@ -447,6 +447,9 @@ vdev_disk_open(vdev_t *v, uint64_t *psize, uint64_t *max_psize,
 	/* Inform the ZIO pipeline that we are non-rotational */
 	v->vdev_nonrot = blk_queue_nonrot(bdev_get_queue(bdev));
 
+	/* Is backed by a block device. */
+	v->vdev_is_blkdev = B_TRUE;
+
 	/* Physical volume size in bytes for the partition */
 	*psize = bdev_capacity(bdev);
 

module/zcommon/zpool_prop.c

@@ -326,6 +326,13 @@ vdev_prop_init(void)
 		{ NULL }
 	};
 
+	static const zprop_index_t vdevschedulertype_table[] = {
+		{ "auto",	VDEV_SCHEDULER_AUTO },
+		{ "on",		VDEV_SCHEDULER_ON },
+		{ "off",	VDEV_SCHEDULER_OFF },
+		{ NULL }
+	};
+
 	struct zfs_mod_supported_features *sfeatures =
 	    zfs_mod_list_supported(ZFS_SYSFS_VDEV_PROPERTIES);
 
@@ -484,6 +491,10 @@ vdev_prop_init(void)
 	zprop_register_index(VDEV_PROP_SLOW_IO_EVENTS, "slow_io_events",
 	    B_TRUE, PROP_DEFAULT, ZFS_TYPE_VDEV, "on | off",
 	    "SLOW_IO_EVENTS", boolean_table, sfeatures);
+	zprop_register_index(VDEV_PROP_SCHEDULER, "scheduler",
+	    VDEV_SCHEDULER_AUTO, PROP_DEFAULT, ZFS_TYPE_VDEV,
+	    "auto | on | off", "IO_SCHEDULER",
+	    vdevschedulertype_table, sfeatures);
 
 	/* hidden properties */
 	zprop_register_hidden(VDEV_PROP_NAME, "name", PROP_TYPE_STRING,

module/zfs/vdev.c

@@ -767,6 +767,8 @@ vdev_alloc_common(spa_t *spa, uint_t id, uint64_t guid, vdev_ops_t *ops)
 	vd->vdev_slow_io_n = vdev_prop_default_numeric(VDEV_PROP_SLOW_IO_N);
 	vd->vdev_slow_io_t = vdev_prop_default_numeric(VDEV_PROP_SLOW_IO_T);
 
+	vd->vdev_scheduler = vdev_prop_default_numeric(VDEV_PROP_SCHEDULER);
+
 	list_link_init(&vd->vdev_config_dirty_node);
 	list_link_init(&vd->vdev_state_dirty_node);
 	list_link_init(&vd->vdev_initialize_node);
@@ -3972,6 +3974,12 @@ vdev_load(vdev_t *vd)
 		if (error && error != ENOENT)
 			vdev_dbgmsg(vd, "vdev_load: zap_lookup(zap=%llu) "
 			    "failed [error=%d]", (u_longlong_t)zapobj, error);
+
+		error = vdev_prop_get_int(vd, VDEV_PROP_SCHEDULER,
+		    &vd->vdev_scheduler);
+		if (error && error != ENOENT)
+			vdev_dbgmsg(vd, "vdev_load: zap_lookup(zap=%llu) "
+			    "failed [error=%d]", (u_longlong_t)zapobj, error);
 	}
 
 	/*
@@ -6259,6 +6267,13 @@ vdev_prop_set(vdev_t *vd, nvlist_t *innvl, nvlist_t *outnvl)
 			}
 			vd->vdev_slow_io_t = intval;
 			break;
+		case VDEV_PROP_SCHEDULER:
+			if (nvpair_value_uint64(elem, &intval) != 0) {
+				error = EINVAL;
+				break;
+			}
+			vd->vdev_scheduler = intval;
+			break;
 		default:
 			/* Most processing is done in vdev_props_set_sync */
 			break;
@@ -6664,6 +6679,7 @@ vdev_prop_get(vdev_t *vd, nvlist_t *innvl, nvlist_t *outnvl)
 			case VDEV_PROP_IO_T:
 			case VDEV_PROP_SLOW_IO_N:
 			case VDEV_PROP_SLOW_IO_T:
+			case VDEV_PROP_SCHEDULER:
 				err = vdev_prop_get_int(vd, prop, &intval);
 				if (err && err != ENOENT)
 					break;

module/zfs/vdev_file.c

@@ -109,6 +109,9 @@ vdev_file_open(vdev_t *vd, uint64_t *psize, uint64_t *max_psize,
 	 */
 	vd->vdev_nonrot = B_TRUE;
 
+	/* Is not backed by a block device. */
+	vd->vdev_is_blkdev = B_FALSE;
+
 	/*
 	 * Allow TRIM on file based vdevs.  This may not always be supported,
 	 * since it depends on your kernel version and underlying filesystem

module/zfs/vdev_queue.c

@@ -879,6 +879,38 @@ again:
 	return (zio);
 }
 
+static boolean_t
+vdev_should_queue_io(zio_t *zio)
+{
+	vdev_t *vd = zio->io_vd;
+	boolean_t should_queue = B_TRUE;
+
+	/*
+	 * Add zio with ZIO_FLAG_NODATA to queue as bypass code
+	 * currently does not handle certain cases (gang abd, raidz
+	 * write aggregation).
+	 */
+	if (zio->io_flags & ZIO_FLAG_NODATA)
+		return (B_TRUE);
+
+	switch (vd->vdev_scheduler) {
+	case VDEV_SCHEDULER_AUTO:
+		if (vd->vdev_nonrot && vd->vdev_is_blkdev)
+			should_queue = B_FALSE;
+		break;
+	case VDEV_SCHEDULER_ON:
+		should_queue = B_TRUE;
+		break;
+	case VDEV_SCHEDULER_OFF:
+		should_queue = B_FALSE;
+		break;
+	default:
+		should_queue = B_TRUE;
+		break;
+	}
+	return (should_queue);
+}
+
 zio_t *
 vdev_queue_io(zio_t *zio)
 {
@@ -922,6 +954,11 @@ vdev_queue_io(zio_t *zio)
 	zio->io_flags |= ZIO_FLAG_DONT_QUEUE;
 	zio->io_timestamp = gethrtime();
 
+	if (!vdev_should_queue_io(zio)) {
+		zio->io_queue_state = ZIO_QS_NONE;
+		return (zio);
+	}
+
 	mutex_enter(&vq->vq_lock);
 	vdev_queue_io_add(vq, zio);
 	nio = vdev_queue_io_to_issue(vq);
@@ -954,6 +991,9 @@ vdev_queue_io_done(zio_t *zio)
 	vq->vq_io_complete_ts = now;
 	vq->vq_io_delta_ts = zio->io_delta = now - zio->io_timestamp;
 
+	if (zio->io_queue_state == ZIO_QS_NONE)
+		return;
+
 	mutex_enter(&vq->vq_lock);
 	vdev_queue_pending_remove(vq, zio);
 

tests/runfiles/common.run

@@ -564,7 +564,7 @@ tags = ['functional', 'cli_root', 'zpool_scrub']
 tests = ['zpool_set_001_pos', 'zpool_set_002_neg', 'zpool_set_003_neg',
     'zpool_set_ashift', 'zpool_set_features', 'vdev_set_001_pos',
     'user_property_001_pos', 'user_property_002_neg',
-    'zpool_set_clear_userprop']
+    'zpool_set_clear_userprop','vdev_set_scheduler']
 tags = ['functional', 'cli_root', 'zpool_set']
 
 [tests/functional/cli_root/zpool_split]

tests/zfs-tests/tests/Makefile.am

@@ -1274,6 +1274,7 @@ nobase_dist_datadir_zfs_tests_tests_SCRIPTS += \
 	functional/cli_root/zpool_set/setup.ksh \
 	functional/cli_root/zpool/setup.ksh \
 	functional/cli_root/zpool_set/vdev_set_001_pos.ksh \
+	functional/cli_root/zpool_set/vdev_set_scheduler.ksh \
 	functional/cli_root/zpool_set/zpool_set_common.kshlib \
 	functional/cli_root/zpool_set/zpool_set_001_pos.ksh \
 	functional/cli_root/zpool_set/zpool_set_002_neg.ksh \

tests/zfs-tests/tests/functional/cli_root/zpool_get/vdev_get.cfg

@@ -77,4 +77,5 @@ typeset -a properties=(
     trim_support
     trim_errors
     slow_ios
+    scheduler
 )

tests/zfs-tests/tests/functional/cli_root/zpool_set/vdev_set_scheduler.ksh

@@ -0,0 +1,93 @@
+#!/bin/ksh -p
+# SPDX-License-Identifier: CDDL-1.0
+#
+# CDDL HEADER START
+#
+# The contents of this file are subject to the terms of the
+# Common Development and Distribution License (the "License").
+# You may not use this file except in compliance with the License.
+#
+# You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
+# or https://opensource.org/licenses/CDDL-1.0.
+# See the License for the specific language governing permissions
+# and limitations under the License.
+#
+# When distributing Covered Code, include this CDDL HEADER in each
+# file and include the License file at usr/src/OPENSOLARIS.LICENSE.
+# If applicable, add the following below this CDDL HEADER, with the
+# fields enclosed by brackets "[]" replaced with your own identifying
+# information: Portions Copyright [yyyy] [name of copyright owner]
+#
+# CDDL HEADER END
+#
+
+#
+# Copyright (c) 2025 by Triad National Security, LLC.
+#
+
+. $STF_SUITE/include/libtest.shlib
+
+#
+# DESCRIPTION:
+# Setting vdev scheduler property while reading from vdev should not cause panic.
+#
+# STRATEGY:
+# 1. Create a zpool
+# 2. Write a file to the pool.
+# 3. Start reading from file, while also setting the scheduler property.
+#
+
+verify_runnable "global"
+
+command -v fio > /dev/null || log_unsupported "fio missing"
+
+function set_scheduler
+{
+	for i in auto on off ; do
+		sleep 0.1
+		zpool set scheduler=$i $TESTPOOL1 $FILEDEV
+	done
+}
+
+function cleanup
+{
+	destroy_pool $TESTPOOL1
+	log_must rm -f $FILEDEV
+}
+
+log_assert "Toggling vdev scheduler property while reading from vdev should not cause panic"
+log_onexit cleanup
+
+# 1. Create a pool
+
+FILEDEV="$TEST_BASE_DIR/filedev.$$"
+log_must truncate -s $(($MINVDEVSIZE * 2)) $FILEDEV
+create_pool $TESTPOOL1 $FILEDEV
+
+mntpnt=$(get_prop mountpoint $TESTPOOL1)
+
+# 2. Write a file to the pool, while also setting the scheduler property.
+
+log_must eval "fio --filename=$mntpnt/foobar --name=write-file \
+		--rw=write --size=$MINVDEVSIZE --bs=128k --numjobs=1 --direct=1 \
+		--ioengine=sync --time_based --runtime=2 &"
+
+ITERATIONS=4
+
+for i in $(seq $ITERATIONS); do
+	log_must set_scheduler
+done;
+wait
+
+# 3. Starting reading from file, while also setting the scheduler property.
+
+log_must eval "fio --filename=$mntpnt/foobar --name=read-file \
+		--rw=read --size=$MINVDEVSIZE --bs=128k --numjobs=1 --direct=1 \
+		--ioengine=sync --time_based --runtime=2 &"
+
+for i in $(seq $ITERATIONS); do
+	log_must set_scheduler
+done;
+wait
+
+log_pass "Setting vdev scheduler property while reading from vdev does not cause panic"