On 30/09/24 12:09, Paul E. McKenney wrote: > On Fri, Sep 13, 2024 at 11:00:39AM -0700, Paul E. McKenney wrote: >> On Fri, Sep 13, 2024 at 06:55:34PM +0200, Valentin Schneider wrote: >> > On 13/09/24 07:08, Paul E. McKenney wrote: >> > > On Sun, Sep 08, 2024 at 09:32:18AM -0700, Paul E. McKenney wrote: >> > >> >> > >> Just following up... >> > >> >> > >> For whatever it is worth, on last night's run of next-20240906, I got >> > >> nine failures out of 100 6-hour runs of rcutorture’s TREE03 scenario. >> > >> These failures were often, but not always, shortly followed by a hard hang. >> > >> >> > >> The warning at line 1995 is the WARN_ON_ONCE(on_dl_rq(dl_se)) >> > >> in enqueue_dl_entity() and the warning at line 1971 is the >> > >> WARN_ON_ONCE(!RB_EMPTY_NODE(&dl_se->rb_node)) in __enqueue_dl_entity(). >> > >> >> > >> The pair of splats is shown below, in case it helps. >> > > >> > > Again following up... >> > > >> > > I am still seeing this on next-20240912, with six failures out of 100 >> > > 6-hour runs of rcutorture’s TREE03 scenario. Statistics suggests that >> > > we not read much into the change in frequency. >> > > >> > > Please let me know if there are any diagnostic patches or options that >> > > I should apply. >> > >> > Hey, sorry I haven't forgotten about this, I've just spread myself a bit >> > too thin and also apparently I'm supposed to prepare some slides for next >> > week, I'll get back to this soonish. >> >> I know that feeling! Just didn't want it to get lost. > > And Peter asked that I send along a reproducer, which I am finally getting > around to doing: > > tools/testing/selftests/rcutorture/bin/kvm.sh --allcpus --duration 12h --configs "100*TREE03" --trust-make > FYI Tomas (on Cc) has been working on getting pretty much this to run on our infra, no hit so far. How much of a pain would it be to record an ftrace trace while this runs? I'm thinking sched_switch, sched_wakeup and function-tracing dl_server_start() and dl_server_stop() would be a start. AIUI this is running under QEMU so we'd need to record the trace within that, I'm guessing we can (ab)use --bootargs to feed it tracing arguments, but how do we get the trace out?