Re: Too many rescheduling interrupts (still!)

[Peter Zijlstra <peterz@infradead.org>]

On Wed, Feb 12, 2014 at 04:59:52PM +0100, Frederic Weisbecker wrote:
> 2014-02-12 11:13 GMT+01:00 Peter Zijlstra <peterz@infradead.org>:
> > On Tue, Feb 11, 2014 at 02:34:11PM -0800, Andy Lutomirski wrote:
> >> On Tue, Feb 11, 2014 at 1:21 PM, Thomas Gleixner <tglx@linutronix.de> wrote:
> >> >> A small number of reschedule interrupts appear to be due to a race:
> >> >> both resched_task and wake_up_idle_cpu do, essentially:
> >> >>
> >> >> set_tsk_need_resched(t);
> >> >> smb_mb();
> >> >> if (!tsk_is_polling(t))
> >> >>   smp_send_reschedule(cpu);
> >> >>
> >> >> The problem is that set_tsk_need_resched wakes the CPU and, if the CPU
> >> >> is too quick (which isn't surprising if it was in C0 or C1), then it
> >> >> could *clear* TS_POLLING before tsk_is_polling is read.
> >
> > Yeah we have the wrong default for the idle loops.. it should default to
> > polling and only switch to !polling at the very last moment if it really
> > needs an interrupt to wake.
> >
> > Changing this requires someone (probably me again :/) to audit all arch
> > cpu idle drivers/functions.
> 
> Looking at wake_up_idle_cpu(), we set need_resched and send the IPI.
> On the other end, the CPU wakes up, exits the idle loop and even goes
> to the scheduler while there is probably no task to schedule.
> 
> I wonder if this is all necessary. All we need is the timer to be
> handled by the dynticks code to re-evaluate the next tick. So calling
> irq_exit() -> tick_nohz_irq_exit() from the scheduler_ipi() should be
> enough.

No no, the idea was to NOT send IPIs. So falling out of idle by writing
TIF_NEED_RESCHED and having the idle loop fixup the timers on its way
back to idle is what you want.