Ensure that http event loop is empty before the loop exit #13485

Close #11777 This also revert #11006, since event_base_loopexit seems to drop unstarted event. So shutdown might take longer.

This should be simple, I don't think libevent would force such workarounds.

since event_base_loopexit seems to drop unstarted event

According to the documentation it ignores pending events but active events are handled.

And looks like sending the reply is an active event ():

https://github.com/bitcoin/bitcoin/blob/d6cf4bd7eb3b675cdffec9884a6033f41033ad82/src/httpserver.cpp#L583-L608 https://github.com/bitcoin/bitcoin/blob/d6cf4bd7eb3b675cdffec9884a6033f41033ad82/src/httpserver.cpp#L519-L525 Maybe there is a problem in when the the event is activ'ated?

Thatevent_active could be called after threadHTTP exited and the event wouldn't get called in rare case. This can be tested by adding sleep(1) between StartShutdown() and return statement. https://github.com/bitcoin/bitcoin/blob/d6cf4bd7eb3b675cdffec9884a6033f41033ad82/src/rpc/server.cpp#L229-L240

@ken2812221 so instead it should make sure the reply is handled before event_base_loopexit is called?

should make sure the reply is handled before threadHTTP loop exit because the loop can exit earlier than event_base_loopexit called. However I have no idea how to do that.

@ken2812221 Are you still working on this?

@MarcoFalke If necessary, I can work on this again.

<!--e57a25ab6845829454e8d69fc972939a-->Reviewers, this pull request conflicts with the following ones:

• #13501 (Correctly terminate HTTP server by promag)

If you consider this pull request important, please also help to review the conflicting pull requests. Ideally, start with the one that should be merged first.

Tested 4e79d3e28167ac169f417623f07f48dd3cb0f6ac on current master and could run the tests for 24 hours in a loop. (Which is not possible on current master)