return 0;
 }
 
-static int writeback_inodes_sb_nr_if_idle_safe(struct super_block *sb,
-                                              unsigned long nr_pages,
-                                              enum wb_reason reason)
+static inline int writeback_inodes_sb_nr_if_idle_safe(struct super_block *sb,
+                                                     unsigned long nr_pages,
+                                                     enum wb_reason reason)
 {
-       if (!writeback_in_progress(sb->s_bdi) &&
-           down_read_trylock(&sb->s_umount)) {
+       /* the flusher is dealing with the dirty inodes now. */
+       if (writeback_in_progress(sb->s_bdi))
+               return 1;
+
+       if (down_read_trylock(&sb->s_umount)) {
                writeback_inodes_sb_nr(sb, nr_pages, reason);
                up_read(&sb->s_umount);
                return 1;
        return 0;
 }
 
+void btrfs_writeback_inodes_sb_nr(struct btrfs_root *root,
+                                 unsigned long nr_pages)
+{
+       struct super_block *sb = root->fs_info->sb;
+       int started;
+
+       /* If we can not start writeback, just sync all the delalloc file. */
+       started = writeback_inodes_sb_nr_if_idle_safe(sb, nr_pages,
+                                                     WB_REASON_FS_FREE_SPACE);
+       if (!started) {
+               /*
+                * We needn't worry the filesystem going from r/w to r/o though
+                * we don't acquire ->s_umount mutex, because the filesystem
+                * should guarantee the delalloc inodes list be empty after
+                * the filesystem is readonly(all dirty pages are written to
+                * the disk).
+                */
+               btrfs_start_delalloc_inodes(root, 0);
+               btrfs_wait_ordered_extents(root, 0);
+       }
+}
+
 /*
  * shrink metadata reservation for delalloc
  */
        while (delalloc_bytes && loops < 3) {
                max_reclaim = min(delalloc_bytes, to_reclaim);
                nr_pages = max_reclaim >> PAGE_CACHE_SHIFT;
-               writeback_inodes_sb_nr_if_idle_safe(root->fs_info->sb,
-                                                   nr_pages,
-                                                   WB_REASON_FS_FREE_SPACE);
-
+               btrfs_writeback_inodes_sb_nr(root, nr_pages);
                /*
                 * We need to wait for the async pages to actually start before
                 * we do anything.