Callback thread safety in 0.6?


#1

Does anything in the recent threading development affect this section of the manual? What I’m most interested in is whether we can pass more complex Julia functions to C libraries that might call them from a different context.

I do a bit of this to handle audio callbacks and unfortunately need to do a little more handling from within the callback (mostly copying buffers that are only accessible within the callback context), so I always have to be really careful there’s zero allocation and get a lot of segfaults during development (or if I turn inlining off). It works pretty reliably now (though there’s an issue on Windows), but the code is pretty nasty and hard to maintain, and feels a bit borderline.


#2

No. <20 characters>.


#3

@ssfrr, could you point to an example of where you do this (do the threadsafe callback thing)? I am trying to decode that section of the manual right now, and could use an actual example from which to work.


#4

My packages probably aren’t the best example, as I’m doing some hacky things to push the data into ringbuffers within the callbacks. I’m actually going to be re-architecting that part pretty soon so that it’s less brittle.

Here’s the general idea though. Say you have a C library libjepsen that lets you register a callback:

void call_me_maybe(int (*callback)(int, void*), void *context)

So it’s expecting a callback of the form:

int handle_callback(int mynumber, void *context)

We’ll assume here that it’s handing you back some kind of data in the mynumber parameter it thinks you might be interested in, as well as whatever context pointer you passed in when you registered the callback.

In Julia you’d use this library with something like this:

cond = AsyncCondition()

# this function CANNOT do anything that interacts with the GC
# note we're assuming that our condition handle is being passed in as context
function callback(number, context)
    ccall(:uv_async_send, Cint, (Ptr{Void}, ), context)
    return 0
end

c_callback = cfunction(callback, Cint, (Cint, Ptr{Void})

# register our callback with the condition handle as the
# context pointer so we can use it from within the callback
ccall((:call_me_maybe, libjepsen),
      Void,
      (Ptr{Void}, Ptr{Void}),
      c_callback,
      cond.handle)

while true
    wait(cond)
    println("this is crazy")
end

Note that if the callback gets called multiple times before the Julia-side scheduler gets around to waking up cond, the while loop will only print once. Also note that this doesn’t provide a good mechanism to pass data between thread contexts. I wrote RingBuffers.jl which lets you read and write from the other thread context, but it makes some assumptions about what’s “safe” to do from the other thread and may be a little dicey. It also needs a documentation update - LMK if you want to use it and I’ll help out.


#5

Thanks for the self-contained example.

One of the things I didn’t know how to deal with was being able to pass data, rather than just executing a thunk. Unwinding julia library code and going though libuv documentation, I figured out that the uv_async_t instance that AsyncCondition.handle points to has a data field that can be set on the C side before uv_async_send is invoked. On the Julia side, uv_handle_data gives you the pointer that was set. It would take some further synchronization machinery to construct an actual blocking function call with a return value.

The use case I have in mind isn’t really the callback problem, but calling into the Julia runtime from an embedding application. I want to be able to invoke Julia from different threads in the application. I believe the problem and its solution are analogous to this, though.

Thanks again.


#6

I had to deal with a usecase where I had to actually pass data along from the callback https://github.com/JuliaGPU/OpenCL.jl/blob/52a6e701183a0324a408db0ae7bb9be00d3037e8/src/event.jl#L97-L140


#7

@vchuravy, just to confirm I’m understanding the code you posted:

  • the callback is good for one invocation only;
  • it doesn’t return a value or block until the “real” code is done, yes?

#8

Yes this version is only good for one invocation only (which is necessary for this particular usecase). I presume that you mean blocking from the C++ side? I tried to implement that once for a different project using libuv mutexes and barriers. If you have control over the C++ side you should be able to create a libuv barrier and pass it to the callback and then wait on it being unlocked from your Julia code. But you won’t be able to reuse your AsyncCallback since those can get lost (if another one fires, before your code get to the previous one)