| Message ID | 20130306111537.492045206@linutronix.de (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
Hi Thomas, On Wed, Mar 06, 2013 at 11:18:35AM +0000, Thomas Gleixner wrote: > Some brilliant hardware implementations wake multiple cores when the > broadcast timer fires. This leads to the following interesting > problem: > > CPU0 CPU1 > wakeup from idle wakeup from idle > > leave broadcast mode leave broadcast mode > restart per cpu timer restart per cpu timer > go back to idle > handle broadcast > (empty mask) > enter broadcast mode > programm broadcast device > enter broadcast mode > programm broadcast device > > So what happens is that due to the forced reprogramming of the cpu > local timer, we need to set a event in the future. Now if we manage to > go back to idle before the timer fires, we switch off the timer and > arm the broadcast device with an already expired time (covered by > forced mode). So in the worst case we repeat the above ping pong > forever. > > Unfortunately we have no information about what caused the wakeup, but > we can check current time against the expiry time of the local cpu. If > the local event is already in the past, we know that the broadcast > timer is about to fire and send an IPI. So we mark ourself as an IPI > target even if we left broadcast mode and avoid the reprogramming of > the local cpu timer. > > This still leaves the possibility that a CPU which is not handling the > broadcast interrupt is going to reach idle again before the IPI > arrives. This can't be solved in the core code and will be handled in > follow up patches. > > Reported-by: Jason Liu <liu.h.jason@gmail.com> > Signed-off-by: Thomas Gleixner <tglx@linutronix.de> > --- > kernel/time/tick-broadcast.c | 59 ++++++++++++++++++++++++++++++++++++++++++- > 1 file changed, 58 insertions(+), 1 deletion(-) > > Index: tip/kernel/time/tick-broadcast.c > =================================================================== > --- tip.orig/kernel/time/tick-broadcast.c > +++ tip/kernel/time/tick-broadcast.c > @@ -393,6 +393,7 @@ int tick_resume_broadcast(void) > > static cpumask_var_t tick_broadcast_oneshot_mask; > static cpumask_var_t tick_broadcast_pending_mask; > +static cpumask_var_t tick_broadcast_force_mask; > > /* > * Exposed for debugging: see timer_list.c > @@ -462,6 +463,10 @@ again: > } > } > > + /* Take care of enforced broadcast requests */ > + cpumask_or(tmpmask, tmpmask, tick_broadcast_force_mask); > + cpumask_clear(tick_broadcast_force_mask); I tested the set and it works fine on a dual cluster big.LITTLE testchip using broadcast timer to manage deep idle cluster states. Just asking a question: the force mask is cleared before sending the timer IPI. Would not be better to clear it after the IPI is sent in tick_do_broadcast(...) ? Can you spot a regression if we do this ? The idle thread checks that mask with irqs disabled, so it is possible that we clear the mask before the CPU has a chance to get the IPI. If we clear the mask after sending the IPI, we are increasing the chances for the idle thread to get it. It is just a further optimization, just asking, thanks. > + > /* > * Wakeup the cpus which have an expired event. > */ > @@ -497,6 +502,7 @@ void tick_broadcast_oneshot_control(unsi > struct clock_event_device *bc, *dev; > struct tick_device *td; > unsigned long flags; > + ktime_t now; > int cpu; > > /* > @@ -524,7 +530,16 @@ void tick_broadcast_oneshot_control(unsi > WARN_ON_ONCE(cpumask_test_cpu(cpu, tick_broadcast_pending_mask)); > if (!cpumask_test_and_set_cpu(cpu, tick_broadcast_oneshot_mask)) { > clockevents_set_mode(dev, CLOCK_EVT_MODE_SHUTDOWN); > - if (dev->next_event.tv64 < bc->next_event.tv64) > + /* > + * We only reprogram the broadcast timer if we > + * did not mark ourself in the force mask and > + * if the cpu local event is earlier than the > + * broadcast event. If the current CPU is in > + * the force mask, then we are going to be > + * woken by the IPI right away. > + */ > + if (!cpumask_test_cpu(cpu, tick_broadcast_force_mask) && Is the test against tick_broadcast_force_mask necessary if we add the check in the idle thread before entering idle ? It does not hurt, agreed, and we'd better leave it there, it is just for my own understanding, thanks a lot. Having said that, on the series: Tested-by: Lorenzo Pieralisi <lorenzo.pieralisi@arm.com> > + dev->next_event.tv64 < bc->next_event.tv64) > tick_broadcast_set_event(dev->next_event, 1); > } > } else { > @@ -545,6 +560,47 @@ void tick_broadcast_oneshot_control(unsi > tick_broadcast_pending_mask)) > goto out; > > + /* > + * If the pending bit is not set, then we are > + * either the CPU handling the broadcast > + * interrupt or we got woken by something else. > + * > + * We are not longer in the broadcast mask, so > + * if the cpu local expiry time is already > + * reached, we would reprogram the cpu local > + * timer with an already expired event. > + * > + * This can lead to a ping-pong when we return > + * to idle and therefor rearm the broadcast > + * timer before the cpu local timer was able > + * to fire. This happens because the forced > + * reprogramming makes sure that the event > + * will happen in the future and depending on > + * the min_delta setting this might be far > + * enough out that the ping-pong starts. > + * > + * If the cpu local next_event has expired > + * then we know that the broadcast timer > + * next_event has expired as well and > + * broadcast is about to be handled. So we > + * avoid reprogramming and enforce that the > + * broadcast handler, which did not run yet, > + * will invoke the cpu local handler. > + * > + * We cannot call the handler directly from > + * here, because we might be in a NOHZ phase > + * and we did not go through the irq_enter() > + * nohz fixups. > + */ > + now = ktime_get(); > + if (dev->next_event.tv64 <= now.tv64) { > + cpumask_set_cpu(cpu, tick_broadcast_force_mask); > + goto out; > + } > + /* > + * We got woken by something else. Reprogram > + * the cpu local timer device. > + */ > tick_program_event(dev->next_event, 1); > } > } > @@ -686,5 +742,6 @@ void tick_broadcast_init(void) > #ifdef CONFIG_TICK_ONESHOT > alloc_cpumask_var(&tick_broadcast_oneshot_mask, GFP_NOWAIT); > alloc_cpumask_var(&tick_broadcast_pending_mask, GFP_NOWAIT); > + alloc_cpumask_var(&tick_broadcast_force_mask, GFP_NOWAIT); > #endif > } > > > -- > To unsubscribe from this list: send the line "unsubscribe linux-kernel" in > the body of a message to majordomo@vger.kernel.org > More majordomo info at http://vger.kernel.org/majordomo-info.html > Please read the FAQ at http://www.tux.org/lkml/ >
On Wed, 13 Mar 2013, Lorenzo Pieralisi wrote: > > + /* Take care of enforced broadcast requests */ > > + cpumask_or(tmpmask, tmpmask, tick_broadcast_force_mask); > > + cpumask_clear(tick_broadcast_force_mask); > > I tested the set and it works fine on a dual cluster big.LITTLE testchip > using broadcast timer to manage deep idle cluster states. > > Just asking a question: the force mask is cleared before sending the > timer IPI. Would not be better to clear it after the IPI is sent in > > tick_do_broadcast(...) ? > > Can you spot a regression if we do this ? The idle thread checks that > mask with irqs disabled, so it is possible that we clear the mask before > the CPU has a chance to get the IPI. If we clear the mask after sending > the IPI, we are increasing the chances for the idle thread to get it. > > It is just a further optimization, just asking, thanks. Need to think about that. > > @@ -524,7 +530,16 @@ void tick_broadcast_oneshot_control(unsi > > WARN_ON_ONCE(cpumask_test_cpu(cpu, tick_broadcast_pending_mask)); > > if (!cpumask_test_and_set_cpu(cpu, tick_broadcast_oneshot_mask)) { > > clockevents_set_mode(dev, CLOCK_EVT_MODE_SHUTDOWN); > > - if (dev->next_event.tv64 < bc->next_event.tv64) > > + /* > > + * We only reprogram the broadcast timer if we > > + * did not mark ourself in the force mask and > > + * if the cpu local event is earlier than the > > + * broadcast event. If the current CPU is in > > + * the force mask, then we are going to be > > + * woken by the IPI right away. > > + */ > > + if (!cpumask_test_cpu(cpu, tick_broadcast_force_mask) && > Is the test against tick_broadcast_force_mask necessary if we add the check > in the idle thread before entering idle ? It does not hurt, agreed, and we'd > better leave it there, it is just for my own understanding, thanks a lot. Well, it's necessary for all archs which do not have the check (yet). > Having said that, on the series: > > Tested-by: Lorenzo Pieralisi <lorenzo.pieralisi@arm.com> Thanks, tglx
Index: tip/kernel/time/tick-broadcast.c =================================================================== --- tip.orig/kernel/time/tick-broadcast.c +++ tip/kernel/time/tick-broadcast.c @@ -393,6 +393,7 @@ int tick_resume_broadcast(void) static cpumask_var_t tick_broadcast_oneshot_mask; static cpumask_var_t tick_broadcast_pending_mask; +static cpumask_var_t tick_broadcast_force_mask; /* * Exposed for debugging: see timer_list.c @@ -462,6 +463,10 @@ again: } } + /* Take care of enforced broadcast requests */ + cpumask_or(tmpmask, tmpmask, tick_broadcast_force_mask); + cpumask_clear(tick_broadcast_force_mask); + /* * Wakeup the cpus which have an expired event. */ @@ -497,6 +502,7 @@ void tick_broadcast_oneshot_control(unsi struct clock_event_device *bc, *dev; struct tick_device *td; unsigned long flags; + ktime_t now; int cpu; /* @@ -524,7 +530,16 @@ void tick_broadcast_oneshot_control(unsi WARN_ON_ONCE(cpumask_test_cpu(cpu, tick_broadcast_pending_mask)); if (!cpumask_test_and_set_cpu(cpu, tick_broadcast_oneshot_mask)) { clockevents_set_mode(dev, CLOCK_EVT_MODE_SHUTDOWN); - if (dev->next_event.tv64 < bc->next_event.tv64) + /* + * We only reprogram the broadcast timer if we + * did not mark ourself in the force mask and + * if the cpu local event is earlier than the + * broadcast event. If the current CPU is in + * the force mask, then we are going to be + * woken by the IPI right away. + */ + if (!cpumask_test_cpu(cpu, tick_broadcast_force_mask) && + dev->next_event.tv64 < bc->next_event.tv64) tick_broadcast_set_event(dev->next_event, 1); } } else { @@ -545,6 +560,47 @@ void tick_broadcast_oneshot_control(unsi tick_broadcast_pending_mask)) goto out; + /* + * If the pending bit is not set, then we are + * either the CPU handling the broadcast + * interrupt or we got woken by something else. + * + * We are not longer in the broadcast mask, so + * if the cpu local expiry time is already + * reached, we would reprogram the cpu local + * timer with an already expired event. + * + * This can lead to a ping-pong when we return + * to idle and therefor rearm the broadcast + * timer before the cpu local timer was able + * to fire. This happens because the forced + * reprogramming makes sure that the event + * will happen in the future and depending on + * the min_delta setting this might be far + * enough out that the ping-pong starts. + * + * If the cpu local next_event has expired + * then we know that the broadcast timer + * next_event has expired as well and + * broadcast is about to be handled. So we + * avoid reprogramming and enforce that the + * broadcast handler, which did not run yet, + * will invoke the cpu local handler. + * + * We cannot call the handler directly from + * here, because we might be in a NOHZ phase + * and we did not go through the irq_enter() + * nohz fixups. + */ + now = ktime_get(); + if (dev->next_event.tv64 <= now.tv64) { + cpumask_set_cpu(cpu, tick_broadcast_force_mask); + goto out; + } + /* + * We got woken by something else. Reprogram + * the cpu local timer device. + */ tick_program_event(dev->next_event, 1); } } @@ -686,5 +742,6 @@ void tick_broadcast_init(void) #ifdef CONFIG_TICK_ONESHOT alloc_cpumask_var(&tick_broadcast_oneshot_mask, GFP_NOWAIT); alloc_cpumask_var(&tick_broadcast_pending_mask, GFP_NOWAIT); + alloc_cpumask_var(&tick_broadcast_force_mask, GFP_NOWAIT); #endif }
Some brilliant hardware implementations wake multiple cores when the broadcast timer fires. This leads to the following interesting problem: CPU0 CPU1 wakeup from idle wakeup from idle leave broadcast mode leave broadcast mode restart per cpu timer restart per cpu timer go back to idle handle broadcast (empty mask) enter broadcast mode programm broadcast device enter broadcast mode programm broadcast device So what happens is that due to the forced reprogramming of the cpu local timer, we need to set a event in the future. Now if we manage to go back to idle before the timer fires, we switch off the timer and arm the broadcast device with an already expired time (covered by forced mode). So in the worst case we repeat the above ping pong forever. Unfortunately we have no information about what caused the wakeup, but we can check current time against the expiry time of the local cpu. If the local event is already in the past, we know that the broadcast timer is about to fire and send an IPI. So we mark ourself as an IPI target even if we left broadcast mode and avoid the reprogramming of the local cpu timer. This still leaves the possibility that a CPU which is not handling the broadcast interrupt is going to reach idle again before the IPI arrives. This can't be solved in the core code and will be handled in follow up patches. Reported-by: Jason Liu <liu.h.jason@gmail.com> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> --- kernel/time/tick-broadcast.c | 59 ++++++++++++++++++++++++++++++++++++++++++- 1 file changed, 58 insertions(+), 1 deletion(-)