On 2017/1/5 7:35, Theodore Ts'o wrote:

So how exactly how did we get into this state?  When we read the inode
into memory, if i_nlink is zero, we declare the file system as
corrupted immediately.

So I assume this is happening the on-disk i_links_count (which is read
into inode->i_nlink) was too low.  So I think the way we should be
handling this is in unlink and rename, before we let i_nlink drop to
zero, we need to check to see if there are other dcache entries
pointing at the inode.  If so, we need to call ext4_error(), and in
the errors=continue case, return EFSCORRUPTED (aka EUCLEAN).

diff --git a/fs/ext4/namei.c b/fs/ext4/namei.c
--- a/fs/ext4/namei.c
+++ b/fs/ext4/namei.c

@@ -3662,6 +3662,11 @@ static int ext4_rename(struct inode *old_dir, struct dentry *old_dentry,
        }

        if (new.inode) {
+               if (new.inode->i_nlink == 0) {
+                       ext4_warning_inode(new.inode, "Removing file '%.*s' with no links",
+                                      new.dentry->d_name.len, new.dentry->d_name.name);
+                       set_nlink(new.inode, 1);
+               }
                ext4_dec_count(handle, new.inode);
                new.inode->i_ctime = ext4_current_time(new.inode);
        }

Because the filesystem have many errors, and the reason of i_nlink becomes zero is unknown,
the on-disk i_links_count was too low may be one reason.  I think we can add i_nlink check in
ext4_rename just like ext4_unlink did, it can avoid inversion under any case.