feat: add RwSpinlock readers-writer lock

[mkroening]

This adds the RwSpinlock implementation from hermit-sync, which has been derived from spin, but

• with separation of UPGRADABLE and EXCLUSIVE,
• with optional exponential backoff,
• with impl RawRwLockRecursive.

Depends on #16 for the backoff feature.
Closes #15

[mkroening]

Rebased on latest master. Would you like more elaborate documentation and examples for these new types?

[phil-opp]

Looks like this still uses the backoff feature instead of the Relax parameter.

[phil-opp]

Looks good, thanks a lot!

Note that I'm not an expert atomic operations, so I'm not 100% certain that all the orderings etc are correct. They looked all reasonable to me, though, and I didn't notice any issues.

It would be great if we could apply some more tooling to catch potential issues at some point (in some future PR). For example, we could run the randomized lock/unlock tests under a data race detector.

[mkroening]

Looks like this still uses the backoff feature instead of the Relax parameter.

Fixed. I forgot during rebase. 😅

It would be great if we could apply some more tooling to catch potential issues at some point (in some future PR). For example, we could run the randomized lock/unlock tests under a data race detector.

Good point. Did you have any specific tool in mind?

I took a look at loom once, but their atomics are not const-constructible, so that clashes with lock_api. We might work around that, but that is probably a larger effort.

[phil-opp]

Thanks!

I took a look at loom once, but their atomics are not const-constructible, so that clashes with lock_api. We might work around that, but that is probably a larger effort.

Yeah, looks like loom will be difficult to use then. I found this related issue: tokio-rs/loom#170

As an alternative, we could try Miri or ThreadSanitizer, which provide less strong guarantees, but should be easier to set up.

[phil-opp]

Thanks for the quick fixes!