On Sat, Dec 31, 2022 at 12:28:10AM -0500, Joel Fernandes wrote: > On Fri, Dec 30, 2022 at 9:14 PM Paul E. McKenney <paulmck@xxxxxxxxxx> wrote: > > > > On Fri, Dec 30, 2022 at 08:42:24PM -0500, Joel Fernandes wrote: > > > Hello, > > > > > > I have been firefighting a hang on 6.0.y stable kernels with > > > rcutorture. It happens mostly consistently when TREE07 is shutting > > > down. > > > > > > It appears that the RCU torture threads are attempted to stop but the > > > shutdown thread, but are constantly awakened by a timer softirq > > > handler in ksoftirqd context. When they wake up, they immediately goto > > > sleep in uninterruptible state until the next time a timer handler > > > wakes them up. It appears the timer softirq is long enough to cause > > > RCU stalls and I see it calling 100s of timer function handlers > > > (call_timer_fn). > > > > > > I am doing some more investigation with trace_printk(s): > > > https://git.kernel.org/pub/scm/linux/kernel/git/jfern/linux.git/commit/?h=stable/trace-hang-6.0.y&id=b779b1e92c97f29333a282ee8f548da02f64de2b > > > > > > Regarding the timer handlers, I was wondering if it is possible that a > > > large number of timer handlers constantly queued can cause RCU stalls > > > due to the timer softirq taking a very long time. That certainly > > > appears to be the case here. Shouldn't the timer softirq also do > > > rcu_softirq_qs() similar to the ksoftirq loop, in case there are too > > > many of them? > > > > That is certainly a good thing to try! > > I am trying something like this just for testing, let's see what happens ;-) > > @@ -1788,9 +1796,14 @@ static inline void __run_timers(struct timer_base *base) > > while (levels--) > expire_timers(base, heads + levels); Might you need to put it inside the body of this "while" loop? > + > + rcu_softirq_qs(); > } > > I guess I am also wondering why the rcu reader does not stop queuing > timers. It is doing schedule_timeout_interruptible() constantly even > though the test is being stopped. As in it is doing schedule_timeout_interruptible() in a loop that does not check torture_must_stop()? That would be bad unless there is some reason the loop cannot just stop. But if the loop cannot just stop, that leads me to believe that there is a bug elsewhere. After all, given that the test is ending, why shouldn't all the loops be able to just stop? (Maybe it is waiting on a buggy child task, for example.) Thanx, Paul
