On Mon, Aug 12, 2019 at 11:02:33PM +0200, Frederic Weisbecker wrote:
> On Fri, Aug 02, 2019 at 08:15:01AM -0700, Paul E. McKenney wrote:
> > The multi_cpu_stop() function relies on the scheduler to gain control from
> > whatever is running on the various online CPUs, including any nohz_full
> > CPUs running long loops in kernel-mode code. Lack of the scheduler-clock
> > interrupt on such CPUs can delay multi_cpu_stop() for several minutes
> > and can also result in RCU CPU stall warnings. This commit therefore
> > causes multi_cpu_stop() to enable the scheduler-clock interrupt on all
> > online CPUs.
> >
> > Signed-off-by: Paul E. McKenney <paulmck@xxxxxxxxxxxxx>
> > ---
> > kernel/stop_machine.c | 9 ++++++++-
> > 1 file changed, 8 insertions(+), 1 deletion(-)
> >
> > diff --git a/kernel/stop_machine.c b/kernel/stop_machine.c
> > index b4f83f7bdf86..a2659f61ed92 100644
> > --- a/kernel/stop_machine.c
> > +++ b/kernel/stop_machine.c
> > @@ -20,6 +20,7 @@
> > #include <linux/smpboot.h>
> > #include <linux/atomic.h>
> > #include <linux/nmi.h>
> > +#include <linux/tick.h>
> > #include <linux/sched/wake_q.h>
> >
> > /*
> > @@ -187,15 +188,19 @@ static int multi_cpu_stop(void *data)
> > {
> > struct multi_stop_data *msdata = data;
> > enum multi_stop_state curstate = MULTI_STOP_NONE;
> > - int cpu = smp_processor_id(), err = 0;
> > + int cpu, err = 0;
> > const struct cpumask *cpumask;
> > unsigned long flags;
> > bool is_active;
> >
> > + for_each_online_cpu(cpu)
> > + tick_nohz_dep_set_cpu(cpu, TICK_DEP_MASK_RCU);
>
> Looks like it's not the right fix but, should you ever need to set an
> all-CPUs (system wide) tick dependency in the future, you can use tick_set_dep().
Indeed, I have dropped this patch, but I now do something similar in
RCU's CPU-hotplug notifiers. Which does have an effect, especially on
the system that isn't subject to the insane-latency cpu_relax().
Plus I am having to put a similar workaround into RCU's quiescent-state
forcing logic.
But how should this really be done?
Isn't there some sort of monitoring of nohz_full CPUs for accounting
purposes? If so, would it make sense for this monitoring to check for
long-duration kernel execution and enable the tick in this case? The
RCU dyntick machinery can be used to remotely detect the long-duration
kernel execution using something like the following:
int nohz_in_kernel_snap = rcu_dynticks_snap_cpu(cpu);
...
if (rcu_dynticks_in_eqs_cpu(cpu, nohz_in_kernel_snap)
nohz_in_kernel_snap = rcu_dynticks_snap_cpu(cpu);
else
/* Turn on the tick! */
I would supply rcu_dynticks_snap_cpu() and rcu_dynticks_in_eqs_cpu(),
which would be simple wrappers around RCU's private rcu_dynticks_snap()
and rcu_dynticks_in_eqs() functions.
Would this make sense as a general solution, or am I missing a corner
case or three?
Thanx, Paul
