Quoting from the upstream bug:

When a connection is abruptly closed due to an error, and there is still some data left to be read (buffer offset vs buffer bytes), connection_threadmain starts to loop. The connection is marked as closed, which prevents the buffer bytes/offset to be updated (e.g. in connection_read_operation()). So the worker thread endlessly tries to read the "data" over and over, but it should just remove the connection from the conn table since it's marked as closed. While it's looping it's repeatedly taking the factory extension lock which creates a lot of contention with the other worker threads. Ultimately, all the worker threads are trying to read data on closed connections, and the server stops accepting new connections.

Upstream bug (includes patch):

https://fedorahosted.org/389/ticket/48412