[Bug 258412] kevent EVFILT_TIMER, EV_DISPATCH lost if event is re-enabled

https://bugs.freebsd.org/bugzilla/show_bug.cgi?id=258412

Mark Johnston <markj@FreeBSD.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |markj@FreeBSD.org
             Status|New                         |Open

--- Comment #2 from Mark Johnston <markj@FreeBSD.org> ---
Thanks for the report and the pointer to the updated libkqueue tests.  I'm
sorry this didn't get any attention sooner.

I ported three new timer tests to our in-tree copy of the libkqueue test suite,
periodic_modify, periodic_to_oneshot, and dispatch.  They found a couple of
kernel bugs unrelated to the one that you described.

Looking at our documentation, it's not obvious to me that FreeBSD's behaviour
in the dispatch test is incorrect.  With a periodic timer, EV_DISPATCH masks
the knote after its state is retrieved by kqueue(), but the timer continues to
fire and we continue to count events.  So when the timer filter is unmasked
with EV_ENABLE | EV_DISPATCH, the knote is immediately activated.  That seems
consistent with my reading of the documentation.

> The above step fails on FreeBSD 13, we get one event with the data field set to 1.

I would expect this to be 2 = 500ms / 200ms.  That itself is a kernel bug which
I have a patch for.

> Adding EV_ADD to the call re-enabling the timer does nothing, as does only passing EV_ENABLE.

In fact, adding EV_ADD fixes the test for me on FreeBSD.  That also seems to
make sense, since EV_ADD causes the timer state to be reset, throwing away
timer firings that occurred while the knote was masked.

So, I'm not sure what to do here.  I spent some time looking at the XNU sources
to better understand what macos is doing but don't quite understand their timer
state machine yet.

-- 
You are receiving this mail because:
You are the assignee for the bug.