Add knob to disable slow io notifications

Introduce a new vdev property `VDEV_PROP_SLOW_IO_REPORTING` that
allows users to disable notifications for slow devices.
This prevents ZED and/or ZFSD from degrading the pool due to slow
I/O.

Signed-off-by: Mariusz Zaborski <[email protected]>

Author: oshogbo

include/sys/fs/zfs.h

@@ -385,6 +385,7 @@ typedef enum {
 	VDEV_PROP_TRIM_SUPPORT,
 	VDEV_PROP_TRIM_ERRORS,
 	VDEV_PROP_SLOW_IOS,
+	VDEV_PROP_SLOW_IO_EVENTS,
 	VDEV_NUM_PROPS
 } vdev_prop_t;
 

include/sys/vdev_impl.h

@@ -464,6 +464,7 @@ struct vdev {
 	uint64_t	vdev_checksum_t;
 	uint64_t	vdev_io_n;
 	uint64_t	vdev_io_t;
+	boolean_t	vdev_slow_io_events;
 	uint64_t	vdev_slow_io_n;
 	uint64_t	vdev_slow_io_t;
 };

lib/libzfs/libzfs.abi

@@ -5930,7 +5930,8 @@
       <enumerator name='VDEV_PROP_TRIM_SUPPORT' value='49'/>
       <enumerator name='VDEV_PROP_TRIM_ERRORS' value='50'/>
       <enumerator name='VDEV_PROP_SLOW_IOS' value='51'/>
-      <enumerator name='VDEV_NUM_PROPS' value='52'/>
+      <enumerator name='VDEV_PROP_SLOW_IO_EVENTS' value='52'/>
+      <enumerator name='VDEV_NUM_PROPS' value='53'/>
     </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

@@ -45,7 +45,7 @@ section, below.
 Every vdev has a set of properties that export statistics about the vdev
 as well as control various behaviors.
 Properties are not inherited from top-level vdevs, with the exception of
-checksum_n, checksum_t, io_n, io_t, slow_io_n, and slow_io_t.
+checksum_n, checksum_t, io_n, io_t, slow_io_events, slow_io_n, and slow_io_t.
 .Pp
 The values of numeric properties can be specified using human-readable suffixes
 .Po for example,
@@ -126,7 +126,8 @@ Indicates if a leaf device supports trim operations.
 .Pp
 The following native properties can be used to change the behavior of a vdev.
 .Bl -tag -width "allocating"
-.It Sy checksum_n , checksum_t , io_n , io_t , slow_io_n , slow_io_t
+.It Sy checksum_n , checksum_t , io_n , io_t , slow_io_events, slow_io_n ,
+.It Sy slow_io_t
 Tune the fault management daemon by specifying checksum/io thresholds of <N>
 errors in <T> seconds, respectively.
 These properties can be set on leaf and top-level vdevs.

module/zcommon/zpool_prop.c

@@ -475,6 +475,9 @@ vdev_prop_init(void)
 	zprop_register_index(VDEV_PROP_FAILFAST, "failfast", B_TRUE,
 	    PROP_DEFAULT, ZFS_TYPE_VDEV, "on | off", "FAILFAST", boolean_table,
 	    sfeatures);
+	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);
 
 	/* hidden properties */
 	zprop_register_hidden(VDEV_PROP_NAME, "name", PROP_TYPE_STRING,

module/zfs/vdev.c

@@ -427,32 +427,53 @@ vdev_get_nparity(vdev_t *vd)
 }
 
 static int
-vdev_prop_get_int(vdev_t *vd, vdev_prop_t prop, uint64_t *value)
+vdev_prop_get_objid(vdev_t *vd, uint64_t *objid)
 {
-	spa_t *spa = vd->vdev_spa;
-	objset_t *mos = spa->spa_meta_objset;
-	uint64_t objid;
-	int err;
 
 	if (vd->vdev_root_zap != 0) {
-		objid = vd->vdev_root_zap;
+		*objid = vd->vdev_root_zap;
 	} else if (vd->vdev_top_zap != 0) {
-		objid = vd->vdev_top_zap;
+		*objid = vd->vdev_top_zap;
 	} else if (vd->vdev_leaf_zap != 0) {
-		objid = vd->vdev_leaf_zap;
+		*objid = vd->vdev_leaf_zap;
 	} else {
 		return (EINVAL);
 	}
 
+	return (0);
+}
+
+static int
+vdev_prop_get_int(vdev_t *vd, vdev_prop_t prop, uint64_t *value)
+{
+	spa_t *spa = vd->vdev_spa;
+	objset_t *mos = spa->spa_meta_objset;
+	uint64_t objid;
+	int err;
+
+	if (vdev_prop_get_objid(vd, &objid) != 0)
+		return (EINVAL);
+
 	err = zap_lookup(mos, objid, vdev_prop_to_name(prop),
 	    sizeof (uint64_t), 1, value);
-
 	if (err == ENOENT)
 		*value = vdev_prop_default_numeric(prop);
 
 	return (err);
 }
 
+static int
+vdev_prop_get_bool(vdev_t *vd, vdev_prop_t prop, boolean_t *bvalue)
+{
+	int err;
+	uint64_t ivalue;
+
+	err = vdev_prop_get_int(vd, prop, &ivalue);
+	*bvalue = ivalue != 1;
+
+	return (err);
+}
+
 /*
  * Get the number of data disks for a top-level vdev.
  */
@@ -714,8 +735,12 @@ vdev_alloc_common(spa_t *spa, uint_t id, uint64_t guid, vdev_ops_t *ops)
 	 */
 	vd->vdev_checksum_n = vdev_prop_default_numeric(VDEV_PROP_CHECKSUM_N);
 	vd->vdev_checksum_t = vdev_prop_default_numeric(VDEV_PROP_CHECKSUM_T);
+
 	vd->vdev_io_n = vdev_prop_default_numeric(VDEV_PROP_IO_N);
 	vd->vdev_io_t = vdev_prop_default_numeric(VDEV_PROP_IO_T);
+
+	vd->vdev_slow_io_events = vdev_prop_default_numeric(
+	    VDEV_PROP_SLOW_IO_EVENTS);
 	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);
 
@@ -3870,6 +3895,11 @@ vdev_load(vdev_t *vd)
 			vdev_dbgmsg(vd, "vdev_load: zap_lookup(zap=%llu) "
 			    "failed [error=%d]", (u_longlong_t)zapobj, error);
 
+		error = vdev_prop_get_bool(vd, VDEV_PROP_SLOW_IO_EVENTS,
+		    &vd->vdev_slow_io_events);
+		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_SLOW_IO_N,
 		    &vd->vdev_slow_io_n);
 		if (error && error != ENOENT)
@@ -5917,15 +5947,8 @@ vdev_props_set_sync(void *arg, dmu_tx_t *tx)
 	/*
 	 * Set vdev property values in the vdev props mos object.
 	 */
-	if (vd->vdev_root_zap != 0) {
-		objid = vd->vdev_root_zap;
-	} else if (vd->vdev_top_zap != 0) {
-		objid = vd->vdev_top_zap;
-	} else if (vd->vdev_leaf_zap != 0) {
-		objid = vd->vdev_leaf_zap;
-	} else {
+	if (vdev_prop_get_objid(vd, &objid) != 0)
 		panic("unexpected vdev type");
-	}
 
 	mutex_enter(&spa->spa_props_lock);
 
@@ -6102,6 +6125,13 @@ vdev_prop_set(vdev_t *vd, nvlist_t *innvl, nvlist_t *outnvl)
 			}
 			vd->vdev_io_t = intval;
 			break;
+		case VDEV_PROP_SLOW_IO_EVENTS:
+			if (nvpair_value_uint64(elem, &intval) != 0) {
+				error = EINVAL;
+				break;
+			}
+			vd->vdev_slow_io_events = intval != 0;
+			break;
 		case VDEV_PROP_SLOW_IO_N:
 			if (nvpair_value_uint64(elem, &intval) != 0) {
 				error = EINVAL;
@@ -6143,6 +6173,7 @@ vdev_prop_get(vdev_t *vd, nvlist_t *innvl, nvlist_t *outnvl)
 	nvpair_t *elem = NULL;
 	nvlist_t *nvprops = NULL;
 	uint64_t intval = 0;
+	boolean_t boolval = 0;
 	char *strval = NULL;
 	const char *propname = NULL;
 	vdev_prop_t prop;
@@ -6156,15 +6187,8 @@ vdev_prop_get(vdev_t *vd, nvlist_t *innvl, nvlist_t *outnvl)
 
 	nvlist_lookup_nvlist(innvl, ZPOOL_VDEV_PROPS_GET_PROPS, &nvprops);
 
-	if (vd->vdev_root_zap != 0) {
-		objid = vd->vdev_root_zap;
-	} else if (vd->vdev_top_zap != 0) {
-		objid = vd->vdev_top_zap;
-	} else if (vd->vdev_leaf_zap != 0) {
-		objid = vd->vdev_leaf_zap;
-	} else {
+	if (vdev_prop_get_objid(vd, &objid) != 0)
 		return (SET_ERROR(EINVAL));
-	}
 	ASSERT(objid != 0);
 
 	mutex_enter(&spa->spa_props_lock);
@@ -6473,6 +6497,18 @@ vdev_prop_get(vdev_t *vd, nvlist_t *innvl, nvlist_t *outnvl)
 				vdev_prop_add_list(outnvl, propname, strval,
 				    intval, src);
 				break;
+			case VDEV_PROP_SLOW_IO_EVENTS:
+				err = vdev_prop_get_bool(vd, prop, &boolval);
+				if (err && err != ENOENT)
+					break;
+
+				src = ZPROP_SRC_LOCAL;
+				if (boolval == vdev_prop_default_numeric(prop))
+					src = ZPROP_SRC_DEFAULT;
+
+				vdev_prop_add_list(outnvl, propname, NULL,
+				    boolval, src);
+				break;
 			case VDEV_PROP_CHECKSUM_N:
 			case VDEV_PROP_CHECKSUM_T:
 			case VDEV_PROP_IO_N:

module/zfs/zfs_fm.c

@@ -223,6 +223,9 @@ vdev_prop_get_inherited(vdev_t *vd, vdev_prop_t prop)
 		case VDEV_PROP_IO_T:
 			propval = vd->vdev_io_t;
 			break;
+		case VDEV_PROP_SLOW_IO_EVENTS:
+			propval = vd->vdev_slow_io_events;
+			break;
 		case VDEV_PROP_SLOW_IO_N:
 			propval = vd->vdev_slow_io_n;
 			break;

module/zfs/zio.c

@@ -5414,9 +5414,12 @@ zio_done(zio_t *zio)
 				zio->io_vd->vdev_stat.vs_slow_ios++;
 				mutex_exit(&zio->io_vd->vdev_stat_lock);
 
-				(void) zfs_ereport_post(FM_EREPORT_ZFS_DELAY,
-				    zio->io_spa, zio->io_vd, &zio->io_bookmark,
-				    zio, 0);
+				if (zio->io_vd->vdev_slow_io_events) {
+					(void) zfs_ereport_post(
+					    FM_EREPORT_ZFS_DELAY,
+					    zio->io_spa, zio->io_vd,
+					    &zio->io_bookmark, zio, 0);
+				}
 			}
 		}
 	}

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

@@ -71,6 +71,7 @@ typeset -a properties=(
     checksum_t
     io_n
     io_t
+    slow_io_events
     slow_io_n
     slow_io_t
     trim_support

tests/zfs-tests/tests/functional/events/zed_slow_io.ksh

@@ -23,6 +23,7 @@
 
 #
 # Copyright (c) 2023, Klara Inc.
+# Copyright (c) 2025, Mariusz Zaborski <[email protected]>
 #
 
 # DESCRIPTION:
@@ -140,8 +141,8 @@ function slow_io_degrade
 {
 	do_setup
 
-	zpool set slow_io_n=5 $TESTPOOL $VDEV
-	zpool set slow_io_t=60 $TESTPOOL $VDEV
+	log_must zpool set slow_io_n=5 $TESTPOOL $VDEV
+	log_must zpool set slow_io_t=60 $TESTPOOL $VDEV
 
 	start_slow_io
 	for i in {1..16}; do
@@ -193,6 +194,44 @@ function slow_io_no_degrade
 	do_clean
 }
 
+# Change slow_io_n, slow_io_t to 5 events in 60 seconds
+# fire more than 5 events. Disable slow io events.
+# Should not degrade.
+function slow_io_degrade_disabled
+{
+	do_setup
+
+	log_must zpool set slow_io_n=5 $TESTPOOL $VDEV
+	log_must zpool set slow_io_t=60 $TESTPOOL $VDEV
+	log_must zpool set slow_io_events=off $TESTPOOL $VDEV
+
+	start_slow_io
+	for i in {1..16}; do
+		dd if=${FILEPATH}$i of=/dev/null count=1 bs=512 2>/dev/null
+		sleep 0.5
+	done
+	stop_slow_io
+	zpool sync
+
+	#
+	# wait 60 seconds to confirm that zfs.delay was not generated.
+	#
+	typeset -i i=0
+	typeset -i events=0
+	while [[ $i -lt 60 ]]; do
+		events=$(zpool events | grep "ereport\.fs\.zfs.delay" | wc -l)
+		i=$((i+1))
+		sleep 1
+	done
+	log_note "$events delay events found"
+
+	[ $events -eq "0" ] || \
+		log_fail "expecting no delay events, found $events"
+
+	log_mustnot wait_vdev_state $TESTPOOL $VDEV "DEGRADED" 45
+	do_clean
+}
+
 log_assert "Test ZED slow io configurability"
 log_onexit cleanup
 
@@ -202,5 +241,6 @@ log_must zed_start
 default_degrade
 slow_io_degrade
 slow_io_no_degrade
+slow_io_degrade_disabled
 
 log_pass "Test ZED slow io configurability"