Merge recent vendor changes and sync code:

1862 incremental zfs receive fails for sparse file > 8PB
3112 ztest does not honor ZFS_DEBUG
3122 zfs destroy filesystem should prefetch blocks
3129 'zpool reopen' restarts resilvers
3130 ztest failure: Assertion failed:
       0 == dmu_objset_destroy(name, B_FALSE) (0x0 == 0x10)

References:
  https://www.illumos.org/issues/1862
  https://www.illumos.org/issues/3112
  https://www.illumos.org/issues/3122
  https://www.illumos.org/issues/3129
  https://www.illumos.org/issues/3130

Obtained from:	illumos (vendor/illumos, vendor/illumos-sys)
MFC after:	2 weeks

1 files changed, 7 insertions, 5 deletions

diff --git a/sys/cddl/contrib/opensolaris/uts/common/fs/zfs/dsl_dir.c b/sys/cddl/contrib/opensolaris/uts/common/fs/zfs/dsl_dir.c
index 43650f3..6396e1c 100644
--- a/sys/cddl/contrib/opensolaris/uts/common/fs/zfs/dsl_dir.c
+++ b/sys/cddl/contrib/opensolaris/uts/common/fs/zfs/dsl_dir.c
@@ -462,12 +462,14 @@ dsl_dir_destroy_check(void *arg1, void *arg2, dmu_tx_t *tx)
 	/*
 	 * There should be exactly two holds, both from
 	 * dsl_dataset_destroy: one on the dd directory, and one on its
-	 * head ds.  Otherwise, someone is trying to lookup something
-	 * inside this dir while we want to destroy it.  The
-	 * config_rwlock ensures that nobody else opens it after we
-	 * check.
+	 * head ds.  If there are more holds, then a concurrent thread is
+	 * performing a lookup inside this dir while we're trying to destroy
+	 * it.  To minimize this possibility, we perform this check only
+	 * in syncing context and fail the operation if we encounter
+	 * additional holds.  The dp_config_rwlock ensures that nobody else
+	 * opens it after we check.
 	 */
-	if (dmu_buf_refcount(dd->dd_dbuf) > 2)
+	if (dmu_tx_is_syncing(tx) && dmu_buf_refcount(dd->dd_dbuf) > 2)
 		return (EBUSY);
 
 	err = zap_count(mos, dd->dd_phys->dd_child_dir_zapobj, &count);