Use reset event in Kestrel heartbeat to quickly stop thread

[JamesNK]

Fixes #47600

Replace Thread.Sleep(1000) with ManualResetEventSlim.

[JamesNK]

@mconnew This PR updates Kestrel heartbeat to exit immediately and stop before Kestrel starts shutting down. That removes the need to use _stopped to suppress error from heartbeat that you added.

Can you take a look and check that your original issue is still solved?

[JamesNK]

@Tratcher I moved heartbeat stop to the start of shutdown. Can you double check this is ok?

[mconnew]

@JamesNK, it looks like it should resolve it as Dispose() won't return until it will no longer call OnHeartbeat. There's nothing preventing a later change bringing the issue back though as OnHeartbeat is internal so callable. In the catch block you could call WaitOne with a timeout of 0. You can then check the true/false value that's immediately returned to know if it's been set and still avoid logging. That would also enable you to keep the unit test you've removed.

Also, why not ManualResetEventSlim? ManualResetEvent allocates a native win32 handle and the Slim variant is all managed.

[JamesNK]

OnHeartbeat is internal so callable

OnHeartbeat is only internal for unit tests. It won't be called after the heartbeat is disposed.

Also, why not ManualResetEventSlim? ManualResetEvent allocates a native win32 handle and the Slim variant is all managed.

I read that slim implementation offers an advantage for short wait times, which doesn't apply here.

[danmoseley]

In .NET Core and .NET 5+, the default spin-waiting duration is short: on the order of 10s of microseconds, depending on platform and processor. If you expect wait times to be much longer than that, you can still use this class instead of ManualResetEvent (perhaps configured with less or no spin-waiting). However, the performance benefit would likely be only marginal.

So it sounds like it doesn't matter that much, but indeed it should avoid creating the OS object unless you explicitly fetch WaitObject on it, as far as I can see.

[JamesNK]

PR updated to use ManualResetEventSlim. Set spin count to zero as we're waiting a long time (1 sec) between heartbeats. Spinning would be wasted CPU.

Added a test for the heartbeat loop. Previously wasn't tested.