On Tue, Mar 24, 2015 at 6:01 PM, Maxim Dounin <mdounin@mdounin.ru> wrote:
If there are good reasons why the termination takes so long - we
may consider adding another iteration.  Otherwise - yes, it'll
remain fixed.

I have a test to see if the module shuts down properly upon receiving SIGTERM.
This test starts up nginx plus a lot of synthetic load in parallel. 
The SIGTERM signal is sent a few seconds after both are up and running, at which point lots of work will have queued up in the module's worker threads.
Without valgrind, the current timespan nginx allows for wrapping up is (more then) enough, but with valgrind, unwinding takes a longer time. Sometimes less, sometimes more than the current limit nginx imposes, which makes the test unreliable.

Does that count as a good reason?