diff --git a/include/sys/zio.h b/include/sys/zio.h
index 3a756949a4..f985b4cf4d 100644
--- a/include/sys/zio.h
+++ b/include/sys/zio.h
@@ -598,7 +598,7 @@ extern zio_t *zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg,
 
 extern int zio_alloc_zil(spa_t *spa, objset_t *os, uint64_t txg,
     blkptr_t *new_bp, uint64_t size, boolean_t *slog);
-extern void zio_flush(zio_t *zio, vdev_t *vd);
+extern void zio_flush(zio_t *zio, vdev_t *vd, boolean_t propagate);
 extern void zio_shrink(zio_t *zio, uint64_t size);
 
 extern int zio_wait(zio_t *zio);
diff --git a/module/os/freebsd/zfs/vdev_geom.c b/module/os/freebsd/zfs/vdev_geom.c
index b7ff1063b0..7aaa42bfb1 100644
--- a/module/os/freebsd/zfs/vdev_geom.c
+++ b/module/os/freebsd/zfs/vdev_geom.c
@@ -1014,21 +1014,6 @@ vdev_geom_io_intr(struct bio *bp)
 		zio->io_error = SET_ERROR(EIO);
 
 	switch (zio->io_error) {
-	case ENOTSUP:
-		/*
-		 * If we get ENOTSUP for BIO_FLUSH or BIO_DELETE we know
-		 * that future attempts will never succeed. In this case
-		 * we set a persistent flag so that we don't bother with
-		 * requests in the future.
-		 */
-		switch (bp->bio_cmd) {
-		case BIO_FLUSH:
-			vd->vdev_nowritecache = B_TRUE;
-			break;
-		case BIO_DELETE:
-			break;
-		}
-		break;
 	case ENXIO:
 		if (!vd->vdev_remove_wanted) {
 			/*
diff --git a/module/os/linux/zfs/vdev_disk.c b/module/os/linux/zfs/vdev_disk.c
index e69c5f3841..2c963bb05c 100644
--- a/module/os/linux/zfs/vdev_disk.c
+++ b/module/os/linux/zfs/vdev_disk.c
@@ -1232,9 +1232,6 @@ BIO_END_IO_PROTO(vdev_disk_io_flush_completion, bio, error)
 	zio->io_error = -error;
 #endif
 
-	if (zio->io_error && (zio->io_error == EOPNOTSUPP))
-		zio->io_vd->vdev_nowritecache = B_TRUE;
-
 	bio_put(bio);
 	ASSERT3S(zio->io_error, >=, 0);
 	if (zio->io_error)
diff --git a/module/zfs/vdev_label.c b/module/zfs/vdev_label.c
index 47346dd5ac..468390b9ac 100644
--- a/module/zfs/vdev_label.c
+++ b/module/zfs/vdev_label.c
@@ -1830,19 +1830,21 @@ vdev_uberblock_sync_list(vdev_t **svd, int svdcount, uberblock_t *ub, int flags)
 
 	for (int v = 0; v < svdcount; v++) {
 		if (vdev_writeable(svd[v])) {
-			zio_flush(zio, svd[v]);
+			zio_flush(zio, svd[v], B_FALSE);
 		}
 	}
 	if (spa->spa_aux_sync_uber) {
 		spa->spa_aux_sync_uber = B_FALSE;
 		for (int v = 0; v < spa->spa_spares.sav_count; v++) {
 			if (vdev_writeable(spa->spa_spares.sav_vdevs[v])) {
-				zio_flush(zio, spa->spa_spares.sav_vdevs[v]);
+				zio_flush(zio, spa->spa_spares.sav_vdevs[v],
+				    B_FALSE);
 			}
 		}
 		for (int v = 0; v < spa->spa_l2cache.sav_count; v++) {
 			if (vdev_writeable(spa->spa_l2cache.sav_vdevs[v])) {
-				zio_flush(zio, spa->spa_l2cache.sav_vdevs[v]);
+				zio_flush(zio, spa->spa_l2cache.sav_vdevs[v],
+				    B_FALSE);
 			}
 		}
 	}
@@ -2007,13 +2009,13 @@ vdev_label_sync_list(spa_t *spa, int l, uint64_t txg, int flags)
 	zio = zio_root(spa, NULL, NULL, flags);
 
 	for (vd = list_head(dl); vd != NULL; vd = list_next(dl, vd))
-		zio_flush(zio, vd);
+		zio_flush(zio, vd, B_FALSE);
 
 	for (int i = 0; i < 2; i++) {
 		if (!sav[i]->sav_label_sync)
 			continue;
 		for (int v = 0; v < sav[i]->sav_count; v++)
-			zio_flush(zio, sav[i]->sav_vdevs[v]);
+			zio_flush(zio, sav[i]->sav_vdevs[v], B_FALSE);
 		if (l == 1)
 			sav[i]->sav_label_sync = B_FALSE;
 	}
@@ -2091,7 +2093,7 @@ retry:
 	for (vdev_t *vd =
 	    txg_list_head(&spa->spa_vdev_txg_list, TXG_CLEAN(txg)); vd != NULL;
 	    vd = txg_list_next(&spa->spa_vdev_txg_list, vd, TXG_CLEAN(txg)))
-		zio_flush(zio, vd);
+		zio_flush(zio, vd, B_FALSE);
 
 	(void) zio_wait(zio);
 
diff --git a/module/zfs/vdev_raidz.c b/module/zfs/vdev_raidz.c
index 15c8b8ca60..187d3908ff 100644
--- a/module/zfs/vdev_raidz.c
+++ b/module/zfs/vdev_raidz.c
@@ -4172,7 +4172,7 @@ io_error_exit:
 		goto io_error_exit;
 	}
 	pio = zio_root(spa, NULL, NULL, 0);
-	zio_flush(pio, raidvd);
+	zio_flush(pio, raidvd, B_FALSE);
 	zio_wait(pio);
 
 	zfs_dbgmsg("reflow: wrote %llu bytes (logical) to scratch area",
@@ -4231,7 +4231,7 @@ overwrite:
 		goto io_error_exit;
 	}
 	pio = zio_root(spa, NULL, NULL, 0);
-	zio_flush(pio, raidvd);
+	zio_flush(pio, raidvd, B_FALSE);
 	zio_wait(pio);
 
 	zfs_dbgmsg("reflow: overwrote %llu bytes (logical) to real location",
@@ -4339,7 +4339,7 @@ vdev_raidz_reflow_copy_scratch(spa_t *spa)
 	}
 	zio_wait(pio);
 	pio = zio_root(spa, NULL, NULL, 0);
-	zio_flush(pio, raidvd);
+	zio_flush(pio, raidvd, B_FALSE);
 	zio_wait(pio);
 
 	zfs_dbgmsg("reflow recovery: overwrote %llu bytes (logical) "
diff --git a/module/zfs/zil.c b/module/zfs/zil.c
index 3983da6aa4..f451c170fc 100644
--- a/module/zfs/zil.c
+++ b/module/zfs/zil.c
@@ -23,6 +23,7 @@
  * Copyright (c) 2011, 2018 by Delphix. All rights reserved.
  * Copyright (c) 2014 Integros [integros.com]
  * Copyright (c) 2018 Datto Inc.
+ * Copyright (c) 2024, Klara, Inc.
  */
 
 /* Portions Copyright 2010 Robert Milkowski */
@@ -1495,12 +1496,6 @@ zil_lwb_flush_vdevs_done(zio_t *zio)
 		 * includes ZIO errors from either this LWB's write or
 		 * flush, as well as any errors from other dependent LWBs
 		 * (e.g. a root LWB ZIO that might be a child of this LWB).
-		 *
-		 * With that said, it's important to note that LWB flush
-		 * errors are not propagated up to the LWB root ZIO.
-		 * This is incorrect behavior, and results in VDEV flush
-		 * errors not being handled correctly here. See the
-		 * comment above the call to "zio_flush" for details.
 		 */
 
 		zcw->zcw_zio_error = zio->io_error;
@@ -1650,17 +1645,8 @@ zil_lwb_write_done(zio_t *zio)
 
 	while ((zv = avl_destroy_nodes(t, &cookie)) != NULL) {
 		vdev_t *vd = vdev_lookup_top(spa, zv->zv_vdev);
-		if (vd != NULL) {
-			/*
-			 * The "ZIO_FLAG_DONT_PROPAGATE" is currently
-			 * always used within "zio_flush". This means,
-			 * any errors when flushing the vdev(s), will
-			 * (unfortunately) not be handled correctly,
-			 * since these "zio_flush" errors will not be
-			 * propagated up to "zil_lwb_flush_vdevs_done".
-			 */
-			zio_flush(lwb->lwb_root_zio, vd);
-		}
+		if (vd != NULL)
+			zio_flush(lwb->lwb_root_zio, vd, B_TRUE);
 		kmem_free(zv, sizeof (*zv));
 	}
 }
diff --git a/module/zfs/zio.c b/module/zfs/zio.c
index 53992931e0..37a33bb1e8 100644
--- a/module/zfs/zio.c
+++ b/module/zfs/zio.c
@@ -1640,10 +1640,10 @@ zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, abd_t *data, uint64_t size,
  * the flushes complete.
  */
 void
-zio_flush(zio_t *pio, vdev_t *vd)
+zio_flush(zio_t *pio, vdev_t *vd, boolean_t propagate)
 {
-	const zio_flag_t flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE |
-	    ZIO_FLAG_DONT_RETRY;
+	const zio_flag_t flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY |
+	    (propagate ? 0 : ZIO_FLAG_DONT_PROPAGATE);
 
 	if (vd->vdev_nowritecache)
 		return;
@@ -1654,7 +1654,7 @@ zio_flush(zio_t *pio, vdev_t *vd)
 		    NULL, ZIO_STAGE_OPEN, ZIO_FLUSH_PIPELINE));
 	} else {
 		for (uint64_t c = 0; c < vd->vdev_children; c++)
-			zio_flush(pio, vd->vdev_child[c]);
+			zio_flush(pio, vd->vdev_child[c], propagate);
 	}
 }
 
@@ -4553,11 +4553,14 @@ zio_vdev_io_assess(zio_t *zio)
 	/*
 	 * If a cache flush returns ENOTSUP or ENOTTY, we know that no future
 	 * attempts will ever succeed. In this case we set a persistent
-	 * boolean flag so that we don't bother with it in the future.
+	 * boolean flag so that we don't bother with it in the future, and
+	 * then we act like the flush succeeded.
 	 */
 	if ((zio->io_error == ENOTSUP || zio->io_error == ENOTTY) &&
-	    zio->io_type == ZIO_TYPE_FLUSH && vd != NULL)
+	    zio->io_type == ZIO_TYPE_FLUSH && vd != NULL) {
 		vd->vdev_nowritecache = B_TRUE;
+		zio->io_error = 0;
+	}
 
 	if (zio->io_error)
 		zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
diff --git a/tests/runfiles/linux.run b/tests/runfiles/linux.run
index 5817e64900..4a3fcd2cbe 100644
--- a/tests/runfiles/linux.run
+++ b/tests/runfiles/linux.run
@@ -124,6 +124,10 @@ tests = ['auto_offline_001_pos', 'auto_online_001_pos', 'auto_online_002_pos',
     'scrub_after_resilver', 'suspend_resume_single', 'zpool_status_-s']
 tags = ['functional', 'fault']
 
+[tests/functional/flush:Linux]
+tests = ['zil_flush_error']
+tags = ['functional', 'flush']
+
 [tests/functional/features/large_dnode:Linux]
 tests = ['large_dnode_002_pos', 'large_dnode_006_pos', 'large_dnode_008_pos']
 tags = ['functional', 'features', 'large_dnode']
diff --git a/tests/test-runner/bin/zts-report.py.in b/tests/test-runner/bin/zts-report.py.in
index 6db10b91de..39b0ccf594 100755
--- a/tests/test-runner/bin/zts-report.py.in
+++ b/tests/test-runner/bin/zts-report.py.in
@@ -379,6 +379,7 @@ if os.environ.get('CI') == 'true':
         'fault/auto_spare_ashift': ['SKIP', ci_reason],
         'fault/auto_spare_shared': ['SKIP', ci_reason],
         'fault/suspend_resume_single': ['SKIP', ci_reason],
+        'flush/zil_flush_error': ['SKIP', ci_reason],
         'procfs/pool_state': ['SKIP', ci_reason],
     })
 
diff --git a/tests/zfs-tests/include/blkdev.shlib b/tests/zfs-tests/include/blkdev.shlib
index 51eff3023e..bd8557c94b 100644
--- a/tests/zfs-tests/include/blkdev.shlib
+++ b/tests/zfs-tests/include/blkdev.shlib
@@ -462,13 +462,16 @@ function unload_scsi_debug
 # Get scsi_debug device name.
 # Returns basename of scsi_debug device (for example "sdb").
 #
-function get_debug_device
+# $1 (optional): Return the first $1 number of SCSI debug device names.
+function get_debug_device #num
 {
+  typeset num=${1:-1}
+
 	for i in {1..10} ; do
-		val=$(lsscsi | awk '/scsi_debug/ {print $6; exit}' | cut -d/ -f3)
+		val=$(lsscsi | awk '/scsi_debug/ {print $6}' | cut -d/ -f3 | head -n$num)
 
 		# lsscsi can take time to settle
-		if [ "$val" != "-" ] ; then
+		if [[ ! "$val" =~ "-" ]] ; then
 			break
 		fi
 		sleep 1
diff --git a/tests/zfs-tests/tests/Makefile.am b/tests/zfs-tests/tests/Makefile.am
index bbeabc6dfb..34a5a0f075 100644
--- a/tests/zfs-tests/tests/Makefile.am
+++ b/tests/zfs-tests/tests/Makefile.am
@@ -1516,6 +1516,9 @@ nobase_dist_datadir_zfs_tests_tests_SCRIPTS += \
 	functional/features/large_dnode/large_dnode_008_pos.ksh \
 	functional/features/large_dnode/large_dnode_009_pos.ksh \
 	functional/features/large_dnode/setup.ksh \
+	functional/flush/cleanup.ksh \
+	functional/flush/zil_flush_error.ksh \
+	functional/flush/setup.ksh \
 	functional/grow/grow_pool_001_pos.ksh \
 	functional/grow/grow_replicas_001_pos.ksh \
 	functional/history/cleanup.ksh \
diff --git a/tests/zfs-tests/tests/functional/flush/cleanup.ksh b/tests/zfs-tests/tests/functional/flush/cleanup.ksh
new file mode 100755
index 0000000000..4eb59574e4
--- /dev/null
+++ b/tests/zfs-tests/tests/functional/flush/cleanup.ksh
@@ -0,0 +1,28 @@
+#!/bin/ksh -p
+#
+# 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) 2024, Klara, Inc.
+#
+
+. $STF_SUITE/include/libtest.shlib
+
+default_cleanup
diff --git a/tests/zfs-tests/tests/functional/flush/setup.ksh b/tests/zfs-tests/tests/functional/flush/setup.ksh
new file mode 100755
index 0000000000..94a3936ce2
--- /dev/null
+++ b/tests/zfs-tests/tests/functional/flush/setup.ksh
@@ -0,0 +1,30 @@
+#!/bin/ksh -p
+#
+# 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) 2024, Klara, Inc.
+#
+
+. $STF_SUITE/include/libtest.shlib
+
+verify_runnable "global"
+
+log_pass
diff --git a/tests/zfs-tests/tests/functional/flush/zil_flush_error.ksh b/tests/zfs-tests/tests/functional/flush/zil_flush_error.ksh
new file mode 100755
index 0000000000..e053c5d3ba
--- /dev/null
+++ b/tests/zfs-tests/tests/functional/flush/zil_flush_error.ksh
@@ -0,0 +1,259 @@
+#!/bin/ksh -p
+#
+# 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) 2024, Klara, Inc.
+#
+
+#
+# This tests that if the ZIL write sequence fails, it corectly falls back and
+# waits until the transaction has fully committed before returning.
+#
+# When this test was written, the ZIL has a flaw - it assumes that if its
+# writes succeed, then the data is definitely on disk and available for reply
+# if the pool fails. It issues a flush immediately after the write, but does
+# not check it is result. If a disk fails after the data has been accepted into
+# the disk cache, but before it can be written to permanent storage, then
+# fsync() will have returned success even though the data is not stored in the
+# ZIL for replay.
+#
+# If the main pool then fails before the transaction can be written, then data
+# is lost, and fsync() returning success was premature.
+#
+# To prove this, we create a pool with a separate log device. We inject two
+# faults:
+#
+# - ZIL writes appear to succeed, but never make it disk
+# - ZIL flushes fail, and return error
+#
+# We then remove the main pool device, and do a write+fsync. This goes to the
+# ZIL, and appears to succeed. When the txg closes, the write will fail, and
+# the pool suspends.
+#
+# Then, we simulate a reboot by copying the content of the pool devices aside.
+# We restore the pool devices, bring it back online, and export it - we don't
+# need it anymore, but we have to clean up properly. Then we restore the copied
+# content and import the pool, in whatever state it was in when it suspended.
+#
+# Finally, we check the content of the file we wrote to. If it matches what we
+# wrote, then the fsync() was correct, and all is well. If it doesn't match,
+# then the flaw is present, and the test fails.
+#
+# We run the test twice: once without the log device injections, one with. The
+# first confirms the expected behaviour of the ZIL - when the pool is imported,
+# the log is replayed. The second fails as above. When the flaw is corrected,
+# both will succeed, and this overall test succeeds.
+#
+
+. $STF_SUITE/include/libtest.shlib
+
+TMPDIR=${TMPDIR:-$TEST_BASE_DIR}
+
+BACKUP_MAIN="$TMPDIR/backup_main"
+BACKUP_LOG="$TMPDIR/backup_log"
+
+LOOP_LOG="$TMPDIR/loop_log"
+
+DATA_FILE="$TMPDIR/data_file"
+
+verify_runnable "global"
+
+function cleanup
+{
+	zinject -c all
+	destroy_pool $TESTPOOL
+	unload_scsi_debug
+	rm -f $BACKUP_MAIN $BACKUP_LOG $DATA_FILE
+}
+
+log_onexit cleanup
+
+log_assert "verify fsync() waits if the ZIL commit fails"
+
+# create 128K of random data, and take its checksum. we do this up front to
+# ensure we don't get messed up by any latency from reading /dev/random or
+# checksumming the file on the pool
+log_must dd if=/dev/random of=$DATA_FILE bs=128K count=1
+typeset sum=$(sha256digest $DATA_FILE)
+
+# create a virtual scsi device with two device nodes. these are backed by the
+# same memory. we do this because we need to be able to take the device offline
+# properly in order to get the pool to suspend; fault injection on a loop
+# device can't do it. once offline, we can use the second node to take a copy
+# of its state.
+load_scsi_debug 100 1 2 1 '512b'
+set -A sd $(get_debug_device 2)
+
+# create a loop device for the log.
+truncate -s 100M $LOOP_LOG
+typeset ld=$(basename $(losetup -f))
+log_must losetup /dev/$ld $LOOP_LOG
+
+# this function runs the entire test sequence. the option decides if faults
+# are injected on the slog device, mimicking the trigger situation that causes
+# the fsync() bug to occur
+function test_fsync
+{
+	typeset -i do_fault_log="$1"
+
+	log_note "setting up test"
+
+	# create the pool. the main data store is on the scsi device, with the
+	# log on a loopback. we bias the ZIL towards to the log device to try
+	# to ensure that fsync() never involves the main device
+	log_must zpool create -f -O logbias=latency $TESTPOOL ${sd[0]} log $ld
+
+	# create the file ahead of time. the ZIL head structure is created on
+	# first use, and does a full txg wait, which we need to avoid
+	log_must dd if=/dev/zero of=/$TESTPOOL/data_file \
+	    bs=128k count=1 conv=fsync
+	log_must zpool sync
+
+	# arrange for writes to the log device to appear to succeed, and
+	# flushes to fail. this simulates a loss of the device between it
+	# accepting the the write into its cache, but before it can be written
+	# out
+	if [[ $do_fault_log != 0 ]] ; then
+		log_note "injecting log device faults"
+		log_must zinject -d $ld -e noop -T write $TESTPOOL
+		log_must zinject -d $ld -e io -T flush $TESTPOOL
+	fi
+
+	# take the main device offline. there is no IO in flight, so ZFS won't
+	# notice immediately
+	log_note "taking main pool offline"
+	log_must eval "echo offline > /sys/block/${sd[0]}/device/state"
+
+	# write out some data, then call fsync(). there are three possible
+	# results:
+	#
+	# - if the bug is present, fsync() will return success, and dd will
+	#   succeed "immediately"; before the pool suspends
+	# - if the bug is fixed, fsync() will block, the pool will suspend, and
+	#   dd will return success after the pool returns to service
+	# - if something else goes wrong, dd will fail; this may happen before
+	#   or after the pool suspends or returns. this shouldn't happen, and
+	#   should abort the test
+	#
+	# we have to put dd in the background, otherwise if it blocks we will
+	# block with it. what we're interested in is whether or not it succeeds
+	# before the pool is suspended. if it does, then we expect that after
+	# the suspended pool is reimported, the data will have been written
+	log_note "running dd in background to write data and call fsync()"
+	dd if=$DATA_FILE of=/$TESTPOOL/data_file bs=128k count=1 conv=fsync &
+	fsync_pid=$!
+
+	# wait for the pool to suspend. this should happen within ~5s, when the
+	# txg sync tries to write the change to the main device
+	log_note "waiting for pool to suspend"
+	typeset -i tries=10
+	until [[ $(cat /proc/spl/kstat/zfs/$TESTPOOL/state) == "SUSPENDED" ]] ; do
+		if ((tries-- == 0)); then
+			log_fail "pool didn't suspend"
+		fi
+		sleep 1
+	done
+
+	# the pool is suspended. see if dd is still present; if it is, then
+	# it's blocked in fsync(), and we have no expectation that the write
+	# made it to disk. if dd has exited, then its return code will tell
+	# us whether fsync() returned success, or it failed for some other
+	# reason
+	typeset -i fsync_success=0
+	if kill -0 $fsync_pid ; then
+		log_note "dd is blocked; fsync() has not returned"
+	else
+		log_note "dd has finished, ensuring it was successful"
+		log_must wait $fsync_pid
+		fsync_success=1
+	fi
+
+	# pool is suspended. if we online the main device right now, it will
+	# retry writing the transaction, which will succed, and everything will
+	# continue as its supposed to. that's the opposite of what we want; we
+	# want to do an import, as if after reboot, to force the pool to try to
+	# replay the ZIL, so we can compare the final result against what
+	# fsync() told us
+	#
+	# however, right now the pool is wedged. we need to get it back online
+	# so we can export it, so we can do the import. so we need to copy the
+	# entire pool state away. for the scsi device, we can do this through
+	# the second device node. for the loopback, we can copy it directly
+	log_note "taking copy of suspended pool"
+	log_must cp /dev/${sd[1]} $BACKUP_MAIN
+	log_must cp /dev/$ld $BACKUP_LOG
+
+	# bring the entire pool back online, by clearing error injections and
+	# restoring the main device. this will unblock anything still waiting
+	# on it, and tidy up all the internals so we can reset it
+	log_note "bringing pool back online"
+	if [[ $do_fault_log != 0 ]] ; then
+		log_must zinject -c all
+	fi
+	log_must eval "echo running > /sys/block/${sd[0]}/device/state"
+	log_must zpool clear $TESTPOOL
+
+	# now the pool is back online. if dd was blocked, it should now
+	# complete successfully. make sure that's true
+	if [[ $fsync_success == 0 ]] ; then
+		log_note "ensuring blocked dd has now finished"
+		log_must wait $fsync_pid
+	fi
+
+	log_note "exporting pool"
+
+	# pool now clean, export it
+	log_must zpool export $TESTPOOL
+
+	log_note "reverting pool to suspended state"
+
+	# restore the pool to the suspended state, mimicking a reboot
+	log_must cp $BACKUP_MAIN /dev/${sd[0]}
+	log_must cp $BACKUP_LOG /dev/$ld
+
+	# import the crashed pool
+	log_must zpool import $TESTPOOL
+
+	# if fsync() succeeded before the pool suspended, then the ZIL should
+	# have replayed properly and the data is now available on the pool
+	#
+	# note that we don't check the alternative; fsync() blocking does not
+	# mean that data _didn't_ make it to disk, just the ZFS never claimed
+	# that it did. in that case we can't know what _should_ be on disk
+	# right now, so can't check
+	if [[ $fsync_success == 1 ]] ; then
+		log_note "fsync() succeeded earlier; checking data was written correctly"
+		typeset newsum=$(sha256digest /$TESTPOOL/data_file)
+		log_must test "$sum" = "$newsum"
+	fi
+
+	log_note "test finished, cleaning up"
+	log_must zpool destroy -f $TESTPOOL
+}
+
+log_note "first run: ZIL succeeds, and repairs the pool at import"
+test_fsync 0
+
+log_note "second run: ZIL commit fails, and falls back to txg sync"
+test_fsync 1
+
+log_pass "fsync() waits if the ZIL commit fails"