`std::sync::mpsc::Sender::send` can block the current thread

[cyexclam]

Location (URL)

https://doc.rust-lang.org/stable/std/sync/mpsc/struct.Sender.html#method.send

Summary

On the documentation for send(), it states "This method will never block the current thread". However, internally it uses SyncWaker which uses a mutex to notify receivers. I believe this should be changed to something akin to "This method will never block the current thread for an unbounded length of time".

I noticed this because I'm running WebAssembly in a browser's main thread, which disallows any sort of blocking. Thus, sending data through Senders extremely rarely spuriously traps. I can work around this, but IMO the documentation should be clearer because this feels like a potential footgun.

An alternative conclusion is that Sender::send shouldn't block at all and that the current behavior is a bug, but this would require a more involved solution.

[Mark-Simulacrum]

I am worried that this implies that any usage of Mutex or similar primitive becomes a "breaking change" in Sender::send (or elsewhere where we might have similar language). I'm not sure we would want to go that far in guarantees made by std.

I guess technically we could also replace the Mutex with a spin lock on wasm, but that seems like a "strictly worse" workaround...

I'm going to go ahead and nominate this for libs-api to see what we think about this guarantee and whether we want to adjust docs or implementation (or both). Historically getting channels right has been pretty thorny, so I'd prefer to avoid tweaking the implementation too much...