Additional limits on hole reporting

Holding the zp->z_rangelock is insufficient to prevent the
dnode from being re-dirtied in all cases.  To avoid looping
indefinately in dmu_offset_next() on files being actively
written only wait once for a dirty dnode to be synced.  If
after waiting it's still dirty don't report the hole.  This
is always safe.

Signed-off-by: Brian Behlendorf <[email protected]>
Issue openzfs#14512

module/zfs/dmu.c

@@ -2127,7 +2127,7 @@ int
 dmu_offset_next(objset_t *os, uint64_t object, boolean_t hole, uint64_t *off)
 {
 	dnode_t *dn;
-	int err;
+	int restarted = 0, err;
 
 restart:
 	err = dnode_hold(os, object, FTAG, &dn);
@@ -2139,19 +2139,23 @@ dmu_offset_next(objset_t *os, uint64_t object, boolean_t hole, uint64_t *off)
 	if (dnode_is_dirty(dn)) {
 		/*
 		 * If the zfs_dmu_offset_next_sync module option is enabled
-		 * then strict hole reporting has been requested.  Dirty
-		 * dnodes must be synced to disk to accurately report all
-		 * holes.  When disabled dirty dnodes are reported to not
-		 * have any holes which is always safe.
-		 *
-		 * When called by zfs_holey_common() the zp->z_rangelock
-		 * is held to prevent zfs_write() and mmap writeback from
-		 * re-dirtying the dnode after txg_wait_synced().
+		 * then hole reporting has been requested.  Dirty dnodes must
+		 * be synced to disk to accurately report holes.  When called
+		 * by zfs_holey_common() the zp->z_rangelock is held to prevent
+		 * zfs_write() and mmap writeback from re-dirtying the dnode
+		 * after txg_wait_synced().  Regardless, if a dnode has been
+		 * dirtied in consecutive txgs by another caller not holding
+		 * the range lock disable hole reporting to avoid looping.
 		 */
 		if (zfs_dmu_offset_next_sync) {
 			rw_exit(&dn->dn_struct_rwlock);
 			dnode_rele(dn, FTAG);
+
+			if (restarted)
+				return (SET_ERROR(EBUSY));
+
 			txg_wait_synced(dmu_objset_pool(os), 0);
+			restarted = 1;
 			goto restart;
 		}