On Thu, 2018-03-22 at 18:09 +0100, Rafael J. Wysocki wrote: > On Thursday, March 22, 2018 5:32:23 PM CET Rik van Riel wrote: > > On Wed, 2018-03-14 at 15:08 +0100, Rafael J. Wysocki wrote: > > > From: Rafael J. Wysocki <[email protected]> > > > > > > If poll_idle() is allowed to spin until need_resched() returns > > > 'true', > > > it may actually spin for a much longer time than expected by the > > > idle > > > governor, since set_tsk_need_resched() is not always called by > > > the > > > timer interrupt handler. If that happens, the CPU may spend much > > > more time than anticipated in the "polling" state. > > > > > > To prevent that from happening, limit the time of the spinning > > > loop > > > in poll_idle(). > > > > > > Suggested-by: Peter Zijlstra <[email protected]> > > > Signed-off-by: Rafael J. Wysocki <[email protected]> > > > > So ... about bisecting that other patch series... > > > > It turned out I had this patch, which looks so > > obviously correct, as patch #1 in my series. > > > > It also turned out that this patch is responsible > > for the entire 5-10% increase in CPU use for the > > memcache style workload. > > > > I wonder if keeping an idle HT thread much busier > > than before slows down its sibling, or something > > like that. > > Uhm, sorry about this.
No worries, this is why we do patch reviews and
tests in the first place.
> Does it improve if you do something like the below on top of it?
That was my next thing to try, after testing just
the idle nohz series by itself :)
I'll push both into the test systems, and will
get back to you when I have answers.
> > Let me go test the nohz idle series by itself,
> > without this patch.
>
> OK
>
> ---
> drivers/cpuidle/poll_state.c | 6 ++++++
> 1 file changed, 6 insertions(+)
>
> Index: linux-pm/drivers/cpuidle/poll_state.c
> ===================================================================
> --- linux-pm.orig/drivers/cpuidle/poll_state.c
> +++ linux-pm/drivers/cpuidle/poll_state.c
> @@ -10,6 +10,7 @@
> #include <linux/sched/idle.h>
>
> #define POLL_IDLE_TIME_LIMIT (TICK_NSEC / 16)
> +#define POLL_IDLE_COUNT 1000
>
> static int __cpuidle poll_idle(struct cpuidle_device *dev,
> struct cpuidle_driver *drv, int
> index)
> @@ -18,9 +19,14 @@ static int __cpuidle poll_idle(struct cp
>
> local_irq_enable();
> if (!current_set_polling_and_test()) {
> + unsigned int loop_count = 0;
> +
> while (!need_resched()) {
> cpu_relax();
> + if (loop_count++ < POLL_IDLE_COUNT)
> + continue;
>
> + loop_count = 0;
> if (local_clock() - time_start >
> POLL_IDLE_TIME_LIMIT)
> break;
> }
>
>
--
All Rights Reversed.
signature.asc
Description: This is a digitally signed message part
