Re: [PATCH v4 1/3] rust: Add `OnceLite` for executing code once

[Alice Ryhl]

On Thu, Dec 5, 2024 at 7:49 AM <jens.korinth@xxxxxxx> wrote:
>
> I'm afraid you lost me. You wrote:
>
> > Using a Once type for this seems like a good idea to me.
>
> and
>
> > One advantage of using a Once type is that we can use core::sync::atomic
> > until we have working LKMM atomics and then we just swap out the Once
> > type without having to modify the warn_once abstractions.
>
> That made sense to me, so I started in this direction. `std::sync::Once`
> has `is_completed` [1], which made particular sense to implement in my
> mind to increase the utility of `OnceLite`.
>
> [1]: https://doc.rust-lang.org/std/sync/struct.Once.html#method.is_completed

The stdlib Once type guarantees that when call_once has returned, then
the closure has finished running *even* if you are not the caller who
is running the closure. It achieves this by having other callers go to
sleep until the main caller completes. If we actually want to mirror
Once, then we should have that logic too, and I really don't think we
want such complicated logic in our pr_warn_once macro.

> > The purpose is to use this for printing once, and printing doesn't need
> > `is_completed`. We can have another helper for other purposes.
>
> Why have `OnceLite` then at all, instead of the hidden Rust macro that was
> proposed initially?

Well, one advantage is that when Boqun manages to add support for LKMM
atomics, we can switch them out without having to modify macro code.
Moving logic out of macro code is usually a good idea. It's also
possible that there are other user-cases of an OnceLite that doesn't
have is_completed.

Alice