Re: [PATCH 5/8] nohz: Restart the tick from irq exit

[Frederic Weisbecker <fweisbec@gmail.com>]

On Sun, Jun 14, 2015 at 03:00:12PM +0530, Preeti U Murthy wrote:
> On 06/12/2015 06:08 PM, Frederic Weisbecker wrote:
> > On Fri, Jun 12, 2015 at 09:32:45AM +0200, Peter Zijlstra wrote:
> >> On Thu, Jun 11, 2015 at 07:36:05PM +0200, Frederic Weisbecker wrote:
> >>> Restart the tick when necessary from the irq exit path. It makes nohz
> >>> full more flexible and allow it to piggyback the tick restart on the
> >>> scheduler IPI in the future instead of sending a dedicated IPI that
> >>> often doubles the scheduler IPI on task wakeup. This will require
> >>> careful review of resched_curr() callers.
> >>
> >> This seems to assume schedule_ipi() callers use irq_exit(), this is
> >> false.
> > 
> > Indeed there will be that too. Note the current patch doesn't yet rely on
> > schedule_ipi(), we are still using the nohz ipis. But introducing the
> > tick restart on irq exit prepares for later piggybacking on scheduler_ipi().
> > 
> > I think this will involve changes on got_nohz_idle_kick(), renamed to
> > got_nohz_kick() and include nohz full related checks to trigger the
> > irq_enter()/exit() pair.
> 
> I maybe saying something obvious here, nevertheless; I am not sure about
> other archs, but atleast on powerpc after handling an interrupt, we will
> call irq_exit() and reevaluate starting of ticks. So in our case even if
> scheduler_ipi() callers do not call irq_exit(), it will be called after
> handling the reschedule interrupt.

scheduler_ipi() takes care of the call to irq_enter() and irq_exit() when
necessary. Which means that the arch low level handler for scheduler_ipi()
shouldn't call these functions. If it does then it's buggy.