On Wed, Jun 04, 2025 at 02:12:11AM +0100, Tingmao Wang wrote:

On 6/4/25 01:55, Al Viro wrote:

quoted

On Wed, Jun 04, 2025 at 01:45:45AM +0100, Tingmao Wang wrote:

quoted

+		rename_seqcount = read_seqbegin(&rename_lock);
+		if (rename_seqcount % 2 == 1) {

Please, describe the condition when that can happen, preferably
along with a reproducer.

My understanding is that when a rename is in progress the seqcount is odd,
is that correct?

If that's the case, then the fs_race_test in patch 2 should act as a
reproducer, since it's constantly moving the directory.

I can add a comment to explain this, thanks for pointing out.

Please, read through the header declaring those primitives and read the
documentation it refers to - it's useful for background.

What's more, look at the area covered by rename_lock - I seriously suspect
that you are greatly overestimating it.