sx_sleep not waking up when timo expires

[Ian Lepore]

On Sun, 2018-07-29 at 21:41 -0700, John-Mark Gurney wrote:
> Ian Lepore wrote this message on Sun, Jul 29, 2018 at 08:35 -0600:
> > 
> > On Sat, 2018-07-28 at 18:01 -0700, John-Mark Gurney wrote:
> > > 
> > > I recently upgraded my router to an Pine A64-LTS board, and have
> > > hit
> > > the same issue as PR 222126[1].  The solution at the end does not
> > > work
> > > for me, as I do not have that line in my loader.conf:
> > > kern.timecounter.smp_tsc_adjust=1
> > > 
> > > I have verified that the wake up does not happen, as I used a
> > > dtrace
> > > script to verify that pf_purge_expired_states is called or not
> > > called..
> > > When I change the timeout, pf will kick the thread and get things
> > > running again, but it has stopped a couple times later...
> > > 
> > > I'm running a recent SNAPSHOT:
> > > FreeBSD gate2.funkthat.com 12.0-CURRENT FreeBSD 12.0-CURRENT #0
> > > r336134: Mon Jul  9 19:20:11 UTC 2018     root at releng3.nyi.freebs
> > > d.org:/usr/obj/usr/src/arm64.aarch64/sys/GENERIC  arm64
> > > 
> > > This is likely reproducable by just starting pf, even in a pass
> > > all
> > > mode, and watching for when the function stops getting
> > > called...  I'll
> > > see if I can't get an extermely minimal config to reproduce it.
> > > 
> > > Any suggestions?
> > > 
> > > [1] https://bugs.freebsd.org/bugzilla/show_bug.cgi?id=222126
> > > 
> > Sounds like
> > 
> >  https://bugs.freebsd.org/bugzilla/show_bug.cgi?id=229644 
> > 
> > which has some patches attached which reduce but don't quite
> > eliminate
> > the occurrances, so nothing has been committed yet. I just ordered
> > a
> > SOPINE board so I can do some hands-on debugging.
> Hmm... The interesting part is that I haven't seen ANY timer
> stability
> issues though...  ntpd is still running w/o troubles... Only ntpd
> logs I have are leap second file related:
> Jul 28 22:36:56 gate2 ntpd[869]: leapsecond file ('/var/db/ntpd.leap-
> seconds.list'): expired less than 214 days ago
> 

ntpd is just a visible victim of the root problem. You found the
problem more directly with dtrace. The patches in that PR fix the
eventtimer, not the timecounter; give it a try.

-- Ian