#[track_caller] for std::sync::Once

[lukaslueg]

rust/library/std/src/sync/once.rs

Lines 390 to 393 in 3227e35

	POISONED if !ignore_poisoning => {
	// Panic to propagate the poison.
	panic!("Once instance has previously been poisoned");
	}

Trying to initialize a poisoned Once does not track it's caller, yielding panic messages like

thread 'null_data' panicked at 'Once instance has previously been poisoned', library/std/src/sync/once.rs:392:21

Can we add #[track_caller] here? I've seen previous PRs regarding other performance-critical methods get rejected due to the overhead (?) track_caller has.

[yaahc]

I think we could add it, we'd just need to run a perf test to make sure there are no obvious regressions, and adding it doesn't mean we won't remove it in the future if people come back and start complaining about the perf impact.

[Mark-Simulacrum]

It is worth noting that it seems at least plausible that this really wants to track the backtrace of the poisoning origin, not so much the (subsequent) access, right? In many cases this will overlap, of course.

Something like #82271 might be worth considering as an expansion of this.

[reez12g]

@rustbot claim

[reez12g]

@lukaslueg

Trying to initialize a poisoned Once does not track it's caller, yielding panic messages like

Could you please give me the sample code for this?
In writing a test in PR for this case, I don't want to misunderstand the situation of this case.