Ref-less parent walk from Landlock (was: Re: [PATCH v3 bpf-next 1/5] namei: Introduce new helper function path_walk_parent())

[Song Liu]

On Sun, Jun 15, 2025 at 5:24 PM Tingmao Wang <m at maowtm.org> wrote:
[...]
> >
> > I would not want it in the first place. But I have a deep seated
> > aversion to exposing two different variants.
>
> Hi Christian, Jan, Song,
>
> I do appreciate your thoughts here and thanks for taking the time to
> explain.  I just have some specific points which I would like you to
> consider:
>
> Taking a step back, maybe the specific designs need a bit more thought,
> but are you at all open to the idea of letting other subsystems take
> advantage of a rcu-based parent walk?

I cannot really speak for VFS folks, but I guess rcu-based parent walk
out of fs/ is not preferred.

> Testing shows that for specific
> cases of a deep directory hierarchy the speedup (for time in Landlock) can
> be almost 60%, and still very significant for the average case. [1]
[...]
> I'm happy to wait till Song's current patch is finished before continuing
> this, but if there is strong objection to two separate APIs, I would
> really appreciate if we can end up in a state where further change to
> implement this is possible.

In v5, path_walk_parent API is not exported. We can easily change it
in the future. Therefore, I don't think we need to rush into a rcu-walk
design before landing path_walk_parent.

Thanks,
Song

[...]