Semaphoreslim, fail faster for timeout 0

[benaadams]

Fail faster for Wait(0) when no counts available.

Don't allocate in the Wait(0) and WaitAsync(0) fail paths.

1M iters Server GC

method	pre (ms)	post (ms)	improvement
Wait(0)	33,788.25	12.28	x 2751.5
WaitAsync(0)	360.79	24.38	x 14.8

Memory allocations, saves 360MB over the 1M ops

[stephentoub]

Please add a comment about why it's safe to be checking this outside of the lock.

[stephentoub]

Have you run the System.Threading tests from corefx with this?

[benaadams]

Just doing it now (within build time definitions of now)

[benaadams]

@stephentoub tested with
msbuild System.Threading.Tests.builds /p:BUILDTOOLS_OVERRIDE_RUNTIME=I:\Work\GitHub\coreclr\bin\Product\Windows_NT.x64.Release

Failure in AsyncMethodNotifications which I assume is unrealated

  Discovering: System.Threading.Tests
  Discovered:  System.Threading.Tests
  Starting:    System.Threading.Tests
     System.Threading.Tests.AsyncLocalTests.AsyncMethodNotifications [FAIL]
        System.TypeLoadException : Could not load type 'YieldAwaiter' from assembly 'System.Threading.Tasks, Version=4.0.12.0, Culture=neutral, PublicKeyToken=b03f5f7f11d50a3a'.
        Stack Trace:
              at System.Threading.Tests.AsyncLocalTests.<>c__DisplayClass7_0.<AsyncMethodNotifications>b__1()
           I:\Work\GitHub\corefx\src\System.Threading\tests\AsyncLocalTests.cs(251,0): at System.Threading.Tests.AsyncLocalTests.Run(Func`1 func)
           I:\Work\GitHub\corefx\src\System.Threading\tests\AsyncLocalTests.cs(275,0): at System.Threading.Tests.AsyncLocalTests.<AsyncMethodNotifications>d__7.MoveNext()
           --- End of stack trace from previous location where exception was thrown ---
              at System.Runtime.ExceptionServices.ExceptionDispatchInfo.Throw()
              at System.Runtime.CompilerServices.TaskAwaiter.HandleNonSuccessAndDebuggerNotification(Task task)
           --- End of stack trace from previous location where exception was thrown ---
              at System.Runtime.ExceptionServices.ExceptionDispatchInfo.Throw()
              at System.Runtime.CompilerServices.TaskAwaiter.HandleNonSuccessAndDebuggerNotification(Task task)
           --- End of stack trace from previous location where exception was thrown ---
              at System.Runtime.ExceptionServices.ExceptionDispatchInfo.Throw()
              at System.Runtime.CompilerServices.TaskAwaiter.HandleNonSuccessAndDebuggerNotification(Task task)
  Finished:    System.Threading.Tests

  === TEST EXECUTION SUMMARY ===
I:\Work\GitHub\corefx\Tools\tests.targets(260,5): warning :    System.Threading.Tests  Total: 167, Errors: 0, Failed: 1, Skipped: 0, Time: 5.414s [I:\Work\GitHub\corefx\src\System.Threading\tests\System.Threading.Tests.csproj]

Also assume I have override correct, it doesn't shout that its using the coreclr repo in a way I can identify

[benaadams]

Fixed spelling of volatile in comments and squashed

[stephentoub]

Thanks, Ben. LGTM.

Have you measured to see what the throughput improvement is? (And verified that indeed there aren't any allocations?

[benaadams]

As aside, a use case example is choice on whether to go async

private bool IsReadyToSend => _outgoingSends.Wait(0);
private Task ReadyToSend => _outgoingSends.WaitAsync();
private bool MaxOutstandingSendsReached => _outgoingSends.CurrentCount == 0;

private Task SendAsync(BufferSpan span, bool endOfMessage)
{
    if (!IsReadyToSend)
    {
        return SendAsyncAwaited(span, endOfMessage);
    }

    var flushSends = endOfMessage || MaxOutstandingSendsReached;

    Send(GetSegmentFromSpan(span), flushSends);

    if (flushSends && !endOfMessage)
    {
        return AwaitReadyToSend();
    }

    return _completedTask;
}

private async Task AwaitReadyToSend()
{
    await ReadyToSend;
}

private async Task SendAsyncAwaited(BufferSpan span, bool endOfMessage)
{
    await ReadyToSend;

    var flushSends = endOfMessage || MaxOutstandingSendsReached;

    Send(GetSegmentFromSpan(span), flushSends);

    if (flushSends && !endOfMessage)
    {
        await ReadyToSend;
    }
}

[benaadams]

@stephentoub the non-async path fared quite badly before (async blocked: code)

1M iters Server GC

method	pre (ms)	post (ms)	improvement
Wait(0)	33,788.25	12.28	x 2751.5
WaitAsync(0)	360.79	24.38	x 14.8

Just testing allocations

[davidfowl]

typo

[benaadams]

Memory allocations, saves 720MB over the 2M ops

Pre

Post

[benaadams]

Looks like its working 😄

[benaadams]

@dotnet-bot test Linux ARM Emulator Cross Debug Build

[stephentoub]

Thanks, @benaadams. LGTM.

[davidfowl]

👏