On Thu, Oct 30, 2014 at 05:54:35PM +0000, Alexander Gordeev wrote:
> On Wed, Oct 29, 2014 at 05:05:24PM -0700, Paul E. McKenney wrote:
> > > > Although it is true that tiny RCU cannot hang a synchronize_rcu()
> > > > grace period, it most certainly can hang a call_rcu() grace period
> > > > in exactly the same way.
> > >
> > > Sorry for being a pain in the neck - just want to make sure I am
> > > following.
> >
> > No worries!
> >
> > > I only see possibility to cause callbacks not being called for "too long"
> > > in case a system has lots of nested interrupts and rcu_idle_enter_common()
> > > is not being called from hardware interrupt context as result. How could
> > > rcu_is_cpu_rrupt_from_idle() help here?
> >
> > Let's start assuming that something in the idle loop posts a callback,
> > and then let me see if I understand your reasoning...
> >
> > 1. The system is idle and stays that way, no runnable tasks.
> >
> > 2. An interrupt occurs. Upon return from interrupt, rcu_irq_exit()
> > is invoked, which calls rcu_idle_enter_common(), which in turn
> > calls rcu_sched_qs(), which does a raise_softirq(RCU_SOFTIRQ).
> >
> > 3. The softirq happens shortly and invokes rcu_process_callbacks(),
> > which invokes __rcu_process_callbacks().
> >
> > 4. So now callbacks can be invoked. At least they can be if
> > ->donetail has been updated. Which it will have been because
> > rcu_sched_qs() invokes rcu_qsctr_help().
>
> Yes, that is exactly my reasoning.
>
> > So your point that rcu_is_cpu_rrupt_from_idle() might be redundant could
> > well be valid -- sorry for being so dismissive earlier.
> >
> > > > > > Now, if you -can- get the userspace-execution indication into
> > > > > > rcu_irq_exit(), this might be of interest. However, it might be
> > > > > > faster
> > > > > > to simply let the scheduling-clock interrupt do the job as it
> > > > > > currently
> > > > > > does, especially for workloads with lots of interrupts.
> > > > > >
> > > > > > Or did you have something else in mind?
> > > > >
> > > > > Nope. I would even leave as is tiny RCU's rcu_is_cpu_rrupt_from_idle()
> > > > > for clarity then ;)
> > > >
> > > > Also to avoid userspace execution from preventing RCU callbacks from
> > > > ever being invoked. ;-)
> > >
> > > Hmm.. Am I missing something else? I did not remove the userspace check
> > > from the scheduling-clock interrupt:
> > >
> > > @@ -250,7 +240,7 @@ void rcu_bh_qs(void)
> > > void rcu_check_callbacks(int cpu, int user)
> > > {
> > > RCU_TRACE(check_cpu_stalls());
> > > - if (user || rcu_is_cpu_rrupt_from_idle())
> > > + if (user)
> > > rcu_sched_qs();
> > > else if (!in_softirq())
> > > rcu_bh_qs();
> >
> > Probably just me being confused. Hopefully so, as shrinking TINY_RCU
> > further will probably be welcome.
>
> Should I resend tiny-only patch?
Please, but also expand the commit log to contain the above reasoning.
> > Have you done any testing of this change?
>
> Just booted to a unicore kernel and dd'ed 1G of /dev/sda to /dev/null.
OK, that is a start. Could I convince you to build a kernel?
Thanx, Paul
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majord...@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
