Julia: Parallelism: Should `wait(::Condition)` take a `Mutex` to unlock/lock?

[NHDaly]

I saw that Base provides a wait(::Condition) function, to add yourself to the Condition's queue:
https://docs.julialang.org/en/v1/stdlib/Distributed/#Base.wait

https://github.com/JuliaLang/julia/blob/d789231e9985537686052db9b2314c0d51656308/base/event.jl#L40-L51

Unless I'm missing some other mechanism that exists for this, I think we also need a way to provide the Condition with a Mutex that it will unlock before sleeping, and re-lock upon awaking.

After the multithreading support is in and tasks can run in parallel, I think this is necessary in order to use Conditions for synchronization.

The pattern I'm describing is the Mesa Monitor, and the wikipedia article covers the required API nicely:
https://en.wikipedia.org/wiki/Monitor_(synchronization)#Condition_variables_2

We can either have the wait function take a mutex argument, as in the above Wikipedia example, or we could provide the mutex in the Condition constructor.

C++ takes the first approach: void wait( std::unique_lock<std::mutex>& lock )
https://en.cppreference.com/w/cpp/thread/condition_variable/wait

And Golang takes the second approach: func NewCond(l Locker) *Cond
https://golang.org/pkg/sync/#Cond.Wait
https://golang.org/pkg/sync/#NewCond

I'm not totally sure about the tradeoffs between those two, but I guess seeing as most of our Concurrency API is based on Go's, it might be good take the decision they made.

---

Here is an example of using the pattern I'm describing:

m = Mutex()  # global
cv = Condition()   # global

function task1()
  lock(m)
  while (!somepred())
    wait(cv, m)  # unlocks m; waits for signal; acquires m
  end
  # ... do stuff ...
  unlock(m)
end
function task2()
  lock(m)
  changepred()
  notify(cv)
  unlock(m)
end

---

I'm happy to open a PR for this as well, but I just wanted to check that I'm not missing something obvious before doing so! :)

[NHDaly]

I'm not totally sure about the tradeoffs between those two, but I guess seeing as most of our Concurrency API is based on Go's, it might be good take the decision they made.

Having the mutex associated with the CV means you can be sure all the Tasks waiting on your queue will acquire the same lock when they awaken, and maybe that can make you more confident about signaling that Condition.

The downside of course is that you _have_ to use the same mutex for all Tasks on the queue. I don't know of any, but maybe there are situations where this forces you to use multiple CVs instead of just one, and can cause code to be overly complex?

[vtjnash]

I'm working on a PR now to implement this. :)

[NHDaly]

My hero! :)

[JeffBezanson]

Yes, this is one of the needed items in the new multithreading system. I also think we should use the Go-style API.