jeslaspravin/copat

This is an C++ 20 cooperative multi tasking library. I used coroutines to achieve cooperative scheduling.

copat's Issues

Workers only JobSystem has higher chance to deadlock when used incorrectly

The deadlock chance when the parallelFor is used from one of the worker threads. The below example shows the issue.

ReturnableAsync<SomeResult> doSomethingAsync(...)
{
    copat::JobSystem *enqFromJs = co_await copat::GetEnqJobSystem{};
    copat::EJobThreadType releaseToThread = enqFromJs->getCurrentThreadType();
    
    // Switch to worker thread
    co_await copat::SwitchJobThreadAwaiter<copat::EJobThreadType::WorkerThreads>{};
    
    // Has a chance to deadlock as there is a random chance the worker job stealing won't pick this thread 
    // and end up waiting here forever.
    copat::parallelFor(&enqFromJs, [](u32 idx){ ... }, 256);
    
    // Release back to the original thread
    co_await copat::SwitchJobSystemThreadAwaiter{ enqFromJs, releaseToThread };    
}

The solution could be either in the CoPaT code base or in the userland.

User land fix

• Yield after copat::diverge but before copat::converge. Cannot use copat::parallelFor.
• Do not dispatch inside the function. Wrap inside another coroutine that co_await's and then resumes.

CoPaT fix

• Work out a heuristics for job stealing such that distribution will always favor the thread idx with jobs in it. Possible but won't be an elegant and performant solution.
• Introduce a new controller or conductor thread that could be optionally enabled by the user and used in this case. Better solution.