| Message ID | 20150119052754.20256.54721.stgit@preeti.in.ibm.com (mailing list archive) |
|---|---|
| State | Not Applicable |
| Headers | show |
On Mon, 19 Jan 2015, Preeti U Murthy wrote: > An idle cpu enters cpu_idle_poll() if it is set in the tick_broadcast_force_mask. > This is so that it does not incur the overhead of entering idle states when it is expected > to be woken up anytime then through a broadcast IPI. The condition that forces an exit out > of the idle polling is the check on setting of the TIF_NEED_RESCHED flag for the idle thread. > > When the broadcast IPI does arrive, it is not guarenteed that the handler sets the > TIF_NEED_RESCHED flag. Hence although the cpu is cleared in the tick_broadcast_force_mask, > it continues to loop in cpu_idle_poll unnecessarily wasting power. Hence exit the idle > poll loop if the tick_broadcast_force_mask gets cleared and enter idle states. > > Of course if the cpu has entered cpu_idle_poll() on being asked to poll explicitly, > it continues to poll till it is asked to reschedule. > > Signed-off-by: Preeti U Murthy <preeti@linux.vnet.ibm.com> > --- > > kernel/sched/idle.c | 3 ++- > 1 file changed, 2 insertions(+), 1 deletion(-) > > diff --git a/kernel/sched/idle.c b/kernel/sched/idle.c > index c47fce7..aaf1c1d 100644 > --- a/kernel/sched/idle.c > +++ b/kernel/sched/idle.c > @@ -47,7 +47,8 @@ static inline int cpu_idle_poll(void) > rcu_idle_enter(); > trace_cpu_idle_rcuidle(0, smp_processor_id()); > local_irq_enable(); > - while (!tif_need_resched()) > + while (!tif_need_resched() && > + (cpu_idle_force_poll || tick_check_broadcast_expired())) You explain the tick_check_broadcast_expired() bit, but what about the cpu_idle_force_poll part? Thanks, tglx
On 01/20/2015 04:51 PM, Thomas Gleixner wrote: > On Mon, 19 Jan 2015, Preeti U Murthy wrote: >> An idle cpu enters cpu_idle_poll() if it is set in the tick_broadcast_force_mask. >> This is so that it does not incur the overhead of entering idle states when it is expected >> to be woken up anytime then through a broadcast IPI. The condition that forces an exit out >> of the idle polling is the check on setting of the TIF_NEED_RESCHED flag for the idle thread. >> >> When the broadcast IPI does arrive, it is not guarenteed that the handler sets the >> TIF_NEED_RESCHED flag. Hence although the cpu is cleared in the tick_broadcast_force_mask, >> it continues to loop in cpu_idle_poll unnecessarily wasting power. Hence exit the idle >> poll loop if the tick_broadcast_force_mask gets cleared and enter idle states. >> >> Of course if the cpu has entered cpu_idle_poll() on being asked to poll explicitly, >> it continues to poll till it is asked to reschedule. >> >> Signed-off-by: Preeti U Murthy <preeti@linux.vnet.ibm.com> >> --- >> >> kernel/sched/idle.c | 3 ++- >> 1 file changed, 2 insertions(+), 1 deletion(-) >> >> diff --git a/kernel/sched/idle.c b/kernel/sched/idle.c >> index c47fce7..aaf1c1d 100644 >> --- a/kernel/sched/idle.c >> +++ b/kernel/sched/idle.c >> @@ -47,7 +47,8 @@ static inline int cpu_idle_poll(void) >> rcu_idle_enter(); >> trace_cpu_idle_rcuidle(0, smp_processor_id()); >> local_irq_enable(); >> - while (!tif_need_resched()) >> + while (!tif_need_resched() && >> + (cpu_idle_force_poll || tick_check_broadcast_expired())) > > You explain the tick_check_broadcast_expired() bit, but what about the > cpu_idle_force_poll part? The last few lines which say "Of course if the cpu has entered cpu_idle_poll() on being asked to poll explicitly, it continues to poll till it is asked to reschedule" explains the cpu_idle_force_poll part. Perhaps I should s/poll explicitly/do cpu_idle_force_poll ? Regards Preeti U Murthy > > Thanks, > > tglx >
On Tue, 20 Jan 2015, Preeti U Murthy wrote: > On 01/20/2015 04:51 PM, Thomas Gleixner wrote: > > On Mon, 19 Jan 2015, Preeti U Murthy wrote: > >> An idle cpu enters cpu_idle_poll() if it is set in the tick_broadcast_force_mask. > >> This is so that it does not incur the overhead of entering idle states when it is expected > >> to be woken up anytime then through a broadcast IPI. The condition that forces an exit out > >> of the idle polling is the check on setting of the TIF_NEED_RESCHED flag for the idle thread. > >> > >> When the broadcast IPI does arrive, it is not guarenteed that the handler sets the > >> TIF_NEED_RESCHED flag. Hence although the cpu is cleared in the tick_broadcast_force_mask, > >> it continues to loop in cpu_idle_poll unnecessarily wasting power. Hence exit the idle > >> poll loop if the tick_broadcast_force_mask gets cleared and enter idle states. > >> > >> Of course if the cpu has entered cpu_idle_poll() on being asked to poll explicitly, > >> it continues to poll till it is asked to reschedule. > >> > >> Signed-off-by: Preeti U Murthy <preeti@linux.vnet.ibm.com> > >> --- > >> > >> kernel/sched/idle.c | 3 ++- > >> 1 file changed, 2 insertions(+), 1 deletion(-) > >> > >> diff --git a/kernel/sched/idle.c b/kernel/sched/idle.c > >> index c47fce7..aaf1c1d 100644 > >> --- a/kernel/sched/idle.c > >> +++ b/kernel/sched/idle.c > >> @@ -47,7 +47,8 @@ static inline int cpu_idle_poll(void) > >> rcu_idle_enter(); > >> trace_cpu_idle_rcuidle(0, smp_processor_id()); > >> local_irq_enable(); > >> - while (!tif_need_resched()) > >> + while (!tif_need_resched() && > >> + (cpu_idle_force_poll || tick_check_broadcast_expired())) > > > > You explain the tick_check_broadcast_expired() bit, but what about the > > cpu_idle_force_poll part? > > The last few lines which say "Of course if the cpu has entered > cpu_idle_poll() on being asked to poll explicitly, it continues to poll > till it is asked to reschedule" explains the cpu_idle_force_poll part. Well, I read it more than once and did not figure it out. The paragraph describes some behaviour. Now I know it's the behaviour before the patch. So maybe something like this: cpu_idle_poll() is entered when cpu_idle_force_poll is set or tick_check_broadcast_expired() returns true. The exit condition from cpu_idle_poll() is tif_need_resched(). But this does not take into account that cpu_idle_force_poll and tick_check_broadcast_expired() can change without setting the resched flag. So a cpu can be caught in cpu_idle_poll() needlessly and thereby wasting power. Add an explicit check for cpu_idle_force_poll and tick_check_broadcast_expired() to the exit condition of cpu_idle_poll() to avoid this. This explains the technical issue without confusing people with IPIs and other completely irrelevant information. Hmm? Thanks, tglx
On 01/21/2015 03:26 PM, Thomas Gleixner wrote: > On Tue, 20 Jan 2015, Preeti U Murthy wrote: >> On 01/20/2015 04:51 PM, Thomas Gleixner wrote: >>> On Mon, 19 Jan 2015, Preeti U Murthy wrote: >>>> An idle cpu enters cpu_idle_poll() if it is set in the tick_broadcast_force_mask. >>>> This is so that it does not incur the overhead of entering idle states when it is expected >>>> to be woken up anytime then through a broadcast IPI. The condition that forces an exit out >>>> of the idle polling is the check on setting of the TIF_NEED_RESCHED flag for the idle thread. >>>> >>>> When the broadcast IPI does arrive, it is not guarenteed that the handler sets the >>>> TIF_NEED_RESCHED flag. Hence although the cpu is cleared in the tick_broadcast_force_mask, >>>> it continues to loop in cpu_idle_poll unnecessarily wasting power. Hence exit the idle >>>> poll loop if the tick_broadcast_force_mask gets cleared and enter idle states. >>>> >>>> Of course if the cpu has entered cpu_idle_poll() on being asked to poll explicitly, >>>> it continues to poll till it is asked to reschedule. >>>> >>>> Signed-off-by: Preeti U Murthy <preeti@linux.vnet.ibm.com> >>>> --- >>>> >>>> kernel/sched/idle.c | 3 ++- >>>> 1 file changed, 2 insertions(+), 1 deletion(-) >>>> >>>> diff --git a/kernel/sched/idle.c b/kernel/sched/idle.c >>>> index c47fce7..aaf1c1d 100644 >>>> --- a/kernel/sched/idle.c >>>> +++ b/kernel/sched/idle.c >>>> @@ -47,7 +47,8 @@ static inline int cpu_idle_poll(void) >>>> rcu_idle_enter(); >>>> trace_cpu_idle_rcuidle(0, smp_processor_id()); >>>> local_irq_enable(); >>>> - while (!tif_need_resched()) >>>> + while (!tif_need_resched() && >>>> + (cpu_idle_force_poll || tick_check_broadcast_expired())) >>> >>> You explain the tick_check_broadcast_expired() bit, but what about the >>> cpu_idle_force_poll part? >> >> The last few lines which say "Of course if the cpu has entered >> cpu_idle_poll() on being asked to poll explicitly, it continues to poll >> till it is asked to reschedule" explains the cpu_idle_force_poll part. > > Well, I read it more than once and did not figure it out. > > The paragraph describes some behaviour. Now I know it's the behaviour > before the patch. So maybe something like this: > > cpu_idle_poll() is entered when cpu_idle_force_poll is set or > tick_check_broadcast_expired() returns true. The exit condition from > cpu_idle_poll() is tif_need_resched(). > > But this does not take into account that cpu_idle_force_poll and > tick_check_broadcast_expired() can change without setting the > resched flag. So a cpu can be caught in cpu_idle_poll() needlessly > and thereby wasting power. > > Add an explicit check for cpu_idle_force_poll and > tick_check_broadcast_expired() to the exit condition of > cpu_idle_poll() to avoid this. > > This explains the technical issue without confusing people with IPIs > and other completely irrelevant information. Hmm? Yep, much simpler, thanks! I will send out the next version with this changelog. Regards Preeti U Murthy > > Thanks, > > tglx > _______________________________________________ > Linuxppc-dev mailing list > Linuxppc-dev@lists.ozlabs.org > https://lists.ozlabs.org/listinfo/linuxppc-dev >
diff --git a/kernel/sched/idle.c b/kernel/sched/idle.c index c47fce7..aaf1c1d 100644 --- a/kernel/sched/idle.c +++ b/kernel/sched/idle.c @@ -47,7 +47,8 @@ static inline int cpu_idle_poll(void) rcu_idle_enter(); trace_cpu_idle_rcuidle(0, smp_processor_id()); local_irq_enable(); - while (!tif_need_resched()) + while (!tif_need_resched() && + (cpu_idle_force_poll || tick_check_broadcast_expired())) cpu_relax(); trace_cpu_idle_rcuidle(PWR_EVENT_EXIT, smp_processor_id()); rcu_idle_exit();
An idle cpu enters cpu_idle_poll() if it is set in the tick_broadcast_force_mask. This is so that it does not incur the overhead of entering idle states when it is expected to be woken up anytime then through a broadcast IPI. The condition that forces an exit out of the idle polling is the check on setting of the TIF_NEED_RESCHED flag for the idle thread. When the broadcast IPI does arrive, it is not guarenteed that the handler sets the TIF_NEED_RESCHED flag. Hence although the cpu is cleared in the tick_broadcast_force_mask, it continues to loop in cpu_idle_poll unnecessarily wasting power. Hence exit the idle poll loop if the tick_broadcast_force_mask gets cleared and enter idle states. Of course if the cpu has entered cpu_idle_poll() on being asked to poll explicitly, it continues to poll till it is asked to reschedule. Signed-off-by: Preeti U Murthy <preeti@linux.vnet.ibm.com> --- kernel/sched/idle.c | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-)