| Message ID | 1455637383-14412-2-git-send-email-daniel.lezcano@linaro.org (mailing list archive) |
|---|---|
| State | Changes Requested, archived |
| Headers | show |
On Tue, 16 Feb 2016, Daniel Lezcano wrote: > Many IRQs are quiet most of the time, or they tend to come in bursts of > fairly equal time intervals within each burst. It is therefore possible > to detect those IRQs with stable intervals and guestimate when the next > IRQ event is most likely to happen. > > Examples of such IRQs may include audio related IRQs where the FIFO size > and/or DMA descriptor size with the sample rate create stable intervals, > block devices during large data transfers, etc. Even network streaming > of multimedia content creates patterns of periodic network interface IRQs > in some cases. > > This patch adds code to compute the mean interval and variance for each IRQ > over a window of time intervals between IRQ events. Those statistics can > be used to assist cpuidle in selecting the most appropriate sleep state > by predicting the most likely time for the next interrupt. > > Signed-off-by: Daniel Lezcano <daniel.lezcano@linaro.org> The math in next_irq_event() is correct even though I think it could be done more simply. But that can be optimized at a later time. Reviewed-by: Nicolas Pitre <nico@linaro.org> > --- > drivers/cpuidle/Kconfig | 9 ++ > kernel/sched/Makefile | 1 + > kernel/sched/idle-sched.c | 280 ++++++++++++++++++++++++++++++++++++++++++++++ > 3 files changed, 290 insertions(+) > create mode 100644 kernel/sched/idle-sched.c > > diff --git a/drivers/cpuidle/Kconfig b/drivers/cpuidle/Kconfig > index 7e48eb5..d64445b 100644 > --- a/drivers/cpuidle/Kconfig > +++ b/drivers/cpuidle/Kconfig > @@ -23,6 +23,15 @@ config CPU_IDLE_GOV_LADDER > config CPU_IDLE_GOV_MENU > bool "Menu governor (for tickless system)" > > +config CPU_IDLE_GOV_SCHED > + bool "Sched idle governor" > + select IRQ_TIMINGS > + help > + Enables an irq timings tracking mechanism to track the wakeup sources > + of the platform. > + > + If you are unsure, it is safe to say N. > + > config DT_IDLE_STATES > bool > > diff --git a/kernel/sched/Makefile b/kernel/sched/Makefile > index 6768797..f7d5a35 100644 > --- a/kernel/sched/Makefile > +++ b/kernel/sched/Makefile > @@ -19,3 +19,4 @@ obj-$(CONFIG_SCHED_AUTOGROUP) += auto_group.o > obj-$(CONFIG_SCHEDSTATS) += stats.o > obj-$(CONFIG_SCHED_DEBUG) += debug.o > obj-$(CONFIG_CGROUP_CPUACCT) += cpuacct.o > +obj-$(CONFIG_CPU_IDLE_GOV_SCHED) += idle-sched.o > diff --git a/kernel/sched/idle-sched.c b/kernel/sched/idle-sched.c > new file mode 100644 > index 0000000..4dc03da > --- /dev/null > +++ b/kernel/sched/idle-sched.c > @@ -0,0 +1,280 @@ > +/* > + * Copyright (C) 2016 Linaro Ltd, Daniel Lezcano <daniel.lezcano@linaro.org> > + * Nicolas Pitre <nicolas.pitre@linaro.org> > + * > + * This program is free software; you can redistribute it and/or modify > + * it under the terms of the GNU General Public License version 2 as > + * published by the Free Software Foundation. > + * > + */ > +#include <linux/cpuidle.h> > +#include <linux/interrupt.h> > +#include <linux/ktime.h> > +#include <linux/tick.h> > + > +/** > + * irqt_mean - compute the average > + * > + * @irqt: the irq timings structure > + * > + * Returns an u32 corresponding to the mean value, or zero if there is > + * no data > + */ > +static inline u32 irqt_mean(struct irq_timings *irqt) > +{ > + return irqt->sum >> IRQ_TIMINGS_SHIFT; > +} > + > +/** > + * irqt_variance - compute the variance > + * > + * @irqt: the irq timings structure > + * > + * Returns an u64 corresponding to the variance, or zero if there is > + * no data > + */ > +static u64 irqt_variance(struct irq_timings *irqt, u32 mean) > +{ > + int i; > + u64 variance = 0; > + > + /* > + * The variance is the sum of the squared difference to the > + * average divided by the number of elements. > + */ > + for (i = 0; i < IRQ_TIMINGS_SIZE; i++) { > + s32 diff = irqt->values[i] - mean; > + variance += (s64)diff * diff; > + } > + > + return variance >> IRQ_TIMINGS_SHIFT; > +} > + > +/** > + * next_irq_event - returns the next irq event > + * > + * Returns a guess estimate remaining time before an interrupt > + * occurs. The type s64 corresponding in a duration in microsecond. > + * > + * The value is always greater or equal to zero. In case, there is no > + * possible prediction, S64_MAX is returned. > + */ > +static s64 next_irq_event(void) > +{ > + struct irq_timings *irqt; > + unsigned int irq = 0; > + u64 variance, now; > + s64 diff, next = 0, min = S64_MAX; > + u32 interval, mean; > + s32 deviation; > + > + now = local_clock(); > + > + /* > + * Search for the earliest expected interruption. > + */ > + rcu_read_lock(); > + while ((irqt = irqtiming_get_next(&irq))) { > + > + /* > + * No values available. > + */ > + if (!irqt->timestamp) > + continue; > + > + /* > + * This interrupt last triggered more than a second ago. > + * It is definitely not predictable for our purpose anymore. > + */ > + if ((now - irqt->timestamp) > NSEC_PER_SEC) > + continue; > + > + /* > + * If the mean value is null, just ignore this wakeup > + * source. > + */ > + mean = irqt_mean(irqt); > + if (!mean) > + continue; > + > + variance = irqt_variance(irqt, mean); > + /* > + * We want to check the last interval is: > + * > + * mean - stddev < interval < mean + stddev > + * > + * That simplifies to: > + * > + * -stddev < interval - mean < stddev > + * > + * abs(interval - mean) < stddev > + * > + * The standard deviation is the sqrt of the variance: > + * > + * abs(interval - mean) < sqrt(variance) > + * > + * and we want to avoid the sqrt, so we square the > + * equation: > + * > + * (interval - mean)^2 < variance > + * > + * So if the latest value of the stats complies with > + * this condition, then the wakeup source is > + * considered predictable and can be used to predict > + * the next event. > + */ > + interval = irqt->values[irqt->w_index]; > + deviation = interval - mean; > + if ((s64)deviation * deviation > variance) > + continue; > + > + /* > + * Let's compute the next event: the wakeup source is > + * considered predictable, we add the average interval > + * time added to the latest interruption event > + * time. Note the timestamp is stored in nsec while > + * the mean time is microsec. > + */ > + next = irqt->timestamp + (mean << 10); > + > + /* > + * If the interrupt is supposed to happen before the > + * minimum time, then it becomes the minimum. > + */ > + if (next < min) > + min = next; > + } > + rcu_read_unlock(); > + > + /* > + * There is no prediction for any interrupt, so we return > + * S64_MAX. > + */ > + if (!next) > + return S64_MAX; > + > + /* > + * At this point, we have our prediction but the caller is > + * expecting the remaining time before the next event, so > + * compute the expected sleep length. > + */ > + diff = min - now; > + > + /* > + * min and now are nanosecond units, let's convert to > + * microsec the result of the diff. > + */ > + diff >>= 10; > + > + /* > + * The result could be negative for different reasons: > + * - the prediction is incorrect but we can't do anything here > + * except assuming the interrupt occured effectively. > + * > + * - the prediction was too near now and expired while we were > + * in this function. > + * > + * In those cases, we return 0. Otherwise we return the expected > + * duration before an interrupt occurs. > + */ > + return diff > 0 ? diff : 0; > +} > + > +/** > + * sched_idle_next_wakeup - Predict the next wakeup on the current cpu > + * > + * The next event on the cpu is based on a statistic approach of the > + * interrupt events and the timer deterministic value. From the timer > + * or the irqs, we return the one expected to occur first. > + * > + * Returns the expected remaining idle time before being woken up by > + * an interruption. > + */ > +s64 sched_idle_next_wakeup(void) > +{ > + s64 next_timer = ktime_to_us(tick_nohz_get_sleep_length()); > + s64 next_irq = next_irq_event(); > + > + return min(next_irq, next_timer); > +} > + > +/** > + * sched_idle - go to idle for a specified amount of time > + * > + * @duration: the idle duration time > + * @latency: the latency constraint > + * > + * Returns 0 on success, < 0 otherwise. > + */ > +int sched_idle(s64 duration, unsigned int latency) > +{ > + struct cpuidle_device *dev = __this_cpu_read(cpuidle_devices); > + struct cpuidle_driver *drv = cpuidle_get_cpu_driver(dev); > + struct cpuidle_state_usage *su; > + struct cpuidle_state *s; > + int i, ret = 0, index = -1; > + > + rcu_idle_enter(); > + > + /* > + * No cpuidle driver is available, let's use the default arch > + * idle function. > + */ > + if (cpuidle_not_available(drv, dev)) > + goto default_idle; > + > + /* > + * Find the idle state with the lowest power while satisfying > + * our constraints. We will save energy if the duration of the > + * idle time is bigger than the target residency which is the > + * break even point. The choice will be modulated by the > + * latency. > + */ > + for (i = 0; i < drv->state_count; i++) { > + > + s = &drv->states[i]; > + > + su = &dev->states_usage[i]; > + > + if (s->disabled || su->disable) > + continue; > + if (s->target_residency > duration) > + continue; > + if (s->exit_latency > latency) > + continue; > + > + index = i; > + } > + > + /* > + * The idle task must be scheduled, it is pointless to go to > + * idle, just re-enable the interrupt and return. > + */ > + if (current_clr_polling_and_test()) { > + local_irq_enable(); > + goto out; > + } > + > + if (index < 0) { > + /* > + * No idle callbacks fulfilled the constraints, jump > + * to the default function like there wasn't any > + * cpuidle driver. > + */ > + goto default_idle; > + } else { > + /* > + * Enter the idle state previously returned by the > + * governor decision. This function will block until > + * an interrupt occurs and will take care of > + * re-enabling the local interrupts > + */ > + return cpuidle_enter(drv, dev, index); > + } > + > +default_idle: > + default_idle_call(); > +out: > + rcu_idle_exit(); > + return ret; > +} > -- > 1.9.1 > > -- To unsubscribe from this list: send the line "unsubscribe linux-pm" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On Tue, Feb 16, 2016 at 5:44 PM, Nicolas Pitre <nicolas.pitre@linaro.org> wrote: > On Tue, 16 Feb 2016, Daniel Lezcano wrote: > >> Many IRQs are quiet most of the time, or they tend to come in bursts of >> fairly equal time intervals within each burst. It is therefore possible >> to detect those IRQs with stable intervals and guestimate when the next >> IRQ event is most likely to happen. >> >> Examples of such IRQs may include audio related IRQs where the FIFO size >> and/or DMA descriptor size with the sample rate create stable intervals, >> block devices during large data transfers, etc. Even network streaming >> of multimedia content creates patterns of periodic network interface IRQs >> in some cases. >> >> This patch adds code to compute the mean interval and variance for each IRQ >> over a window of time intervals between IRQ events. Those statistics can >> be used to assist cpuidle in selecting the most appropriate sleep state >> by predicting the most likely time for the next interrupt. >> >> Signed-off-by: Daniel Lezcano <daniel.lezcano@linaro.org> > > The math in next_irq_event() is correct even though I think it could be > done more simply. But that can be optimized at a later time. > > Reviewed-by: Nicolas Pitre <nico@linaro.org> Well, I'm likely overlooking something, but how is this going to be hooked up to the code in idle.c? Thanks, Rafael -- To unsubscribe from this list: send the line "unsubscribe linux-pm" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On Thu, Feb 18, 2016 at 12:09 AM, Rafael J. Wysocki <rafael@kernel.org> wrote: > On Tue, Feb 16, 2016 at 5:44 PM, Nicolas Pitre <nicolas.pitre@linaro.org> wrote: >> On Tue, 16 Feb 2016, Daniel Lezcano wrote: >> >>> Many IRQs are quiet most of the time, or they tend to come in bursts of >>> fairly equal time intervals within each burst. It is therefore possible >>> to detect those IRQs with stable intervals and guestimate when the next >>> IRQ event is most likely to happen. >>> >>> Examples of such IRQs may include audio related IRQs where the FIFO size >>> and/or DMA descriptor size with the sample rate create stable intervals, >>> block devices during large data transfers, etc. Even network streaming >>> of multimedia content creates patterns of periodic network interface IRQs >>> in some cases. >>> >>> This patch adds code to compute the mean interval and variance for each IRQ >>> over a window of time intervals between IRQ events. Those statistics can >>> be used to assist cpuidle in selecting the most appropriate sleep state >>> by predicting the most likely time for the next interrupt. >>> >>> Signed-off-by: Daniel Lezcano <daniel.lezcano@linaro.org> >> >> The math in next_irq_event() is correct even though I think it could be >> done more simply. But that can be optimized at a later time. >> >> Reviewed-by: Nicolas Pitre <nico@linaro.org> > > Well, I'm likely overlooking something, but how is this going to be > hooked up to the code in idle.c? My somewhat educated guess is that sched_idle() in your patch is intended to replace cpuidle_idle_call(), right? If so, why do you want to replace it? And assuming that you have a good enough reason to do that, you need to ensure that suspend-to-idle will work anyway. Thanks, Rafael -- To unsubscribe from this list: send the line "unsubscribe linux-pm" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On 02/17/2016 11:21 PM, Rafael J. Wysocki wrote: [ ... ] >>> Reviewed-by: Nicolas Pitre <nico@linaro.org> >> >> Well, I'm likely overlooking something, but how is this going to be >> hooked up to the code in idle.c? > > My somewhat educated guess is that sched_idle() in your patch is > intended to replace cpuidle_idle_call(), right? Well, no. I was planning to first have it to use a different code path as experimental code in order to focus improving the accuracy of the prediction and then merge or replace cpuidle_idle_call() with sched_idle(). > If so, why do you want to replace it? > > And assuming that you have a good enough reason to do that, you need > to ensure that suspend-to-idle will work anyway. Yes, sure.
On Thu, Feb 18, 2016 at 11:25 AM, Daniel Lezcano <daniel.lezcano@linaro.org> wrote: > On 02/17/2016 11:21 PM, Rafael J. Wysocki wrote: > > [ ... ] > >>>> Reviewed-by: Nicolas Pitre <nico@linaro.org> >>> >>> >>> Well, I'm likely overlooking something, but how is this going to be >>> hooked up to the code in idle.c? >> >> >> My somewhat educated guess is that sched_idle() in your patch is >> intended to replace cpuidle_idle_call(), right? > > > Well, no. I was planning to first have it to use a different code path as > experimental code in order to focus improving the accuracy of the prediction > and then merge or replace cpuidle_idle_call() with sched_idle(). In that case, what about making it a proper cpuidle governor that people can test and play with in a usual way? Then it may potentially benefit everybody and not just your experimental setup and you may get coverage on systems you have no access to normally. There is some boilerplate code to add for this purpose, but that's not that bad IMO. > >> If so, why do you want to replace it? >> So I'm still unsure why you want to replace cpuidle_idle_call() with sched_idle(). Is there anything wrong with it that it needs to be replaced? Thanks, Rafael -- To unsubscribe from this list: send the line "unsubscribe linux-pm" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On 02/18/2016 07:57 PM, Rafael J. Wysocki wrote: > On Thu, Feb 18, 2016 at 11:25 AM, Daniel Lezcano > <daniel.lezcano@linaro.org> wrote: >> On 02/17/2016 11:21 PM, Rafael J. Wysocki wrote: >> >> [ ... ] >> >>>>> Reviewed-by: Nicolas Pitre <nico@linaro.org> >>>> >>>> >>>> Well, I'm likely overlooking something, but how is this going to be >>>> hooked up to the code in idle.c? >>> >>> >>> My somewhat educated guess is that sched_idle() in your patch is >>> intended to replace cpuidle_idle_call(), right? >> >> >> Well, no. I was planning to first have it to use a different code path as >> experimental code in order to focus improving the accuracy of the prediction >> and then merge or replace cpuidle_idle_call() with sched_idle(). > > In that case, what about making it a proper cpuidle governor that > people can test and play with in a usual way? Then it may potentially > benefit everybody and not just your experimental setup and you may get > coverage on systems you have no access to normally. > > There is some boilerplate code to add for this purpose, but that's not > that bad IMO. Hi Rafael, sorry for the delay in the responses. Actually, adding a new governor is precisely what I would like to avoid because the objective is the scheduler acts as the governor. Here, it is the 'predictor' and the API to enter an idle state conforming the idle duration and the latency constraint. Concerning the testing, it is quite easy to switch from idle_sched to 'menu' via on sched_debug or whatever option we want to add. > > So I'm still unsure why you want to replace cpuidle_idle_call() with > sched_idle(). Is there anything wrong with it that it needs to be > replaced? I don't want to replace cpuidle_idle_call() with sched_idle(). How we integrate the API is something I would like to discuss with another patchset focused in this integration only. For reference: https://lkml.org/lkml/2016/1/12/530 -- Daniel
On Fri, Feb 19, 2016 at 4:01 PM, Daniel Lezcano <daniel.lezcano@linaro.org> wrote: > On 02/18/2016 07:57 PM, Rafael J. Wysocki wrote: >> >> On Thu, Feb 18, 2016 at 11:25 AM, Daniel Lezcano >> <daniel.lezcano@linaro.org> wrote: >>> >>> On 02/17/2016 11:21 PM, Rafael J. Wysocki wrote: >>> >>> [ ... ] >>> >>>>>> Reviewed-by: Nicolas Pitre <nico@linaro.org> >>>>> >>>>> >>>>> >>>>> Well, I'm likely overlooking something, but how is this going to be >>>>> hooked up to the code in idle.c? >>>> >>>> >>>> >>>> My somewhat educated guess is that sched_idle() in your patch is >>>> intended to replace cpuidle_idle_call(), right? >>> >>> >>> >>> Well, no. I was planning to first have it to use a different code path as >>> experimental code in order to focus improving the accuracy of the >>> prediction >>> and then merge or replace cpuidle_idle_call() with sched_idle(). >> >> >> In that case, what about making it a proper cpuidle governor that >> people can test and play with in a usual way? Then it may potentially >> benefit everybody and not just your experimental setup and you may get >> coverage on systems you have no access to normally. >> >> There is some boilerplate code to add for this purpose, but that's not >> that bad IMO. > > > Hi Rafael, > > sorry for the delay in the responses. > > Actually, adding a new governor is precisely what I would like to avoid > because the objective is the scheduler acts as the governor. But why, really? Well, first of all I'm not sure what "the scheduler acts as the governor" means. For the lack of a better explanation I'll refer to the message at https://lkml.org/lkml/2016/1/12/530 that you pointed me at. Your code in there does something like: if (sched_idle_enabled()) { int latency = pm_qos_request(PM_QOS_CPU_DMA_LATENCY); s64 duration = sched_idle_next_wakeup(); sched_idle(duration, latency); } else { cpuidle_idle_call(); } which is quite questionable to be honest as it adds an extra branch to the idle loop for no real benefit. Now, what really is the difference between "governor" and "predictor"? I don't quite see it except that the former is expected to provide a specific interface. The way the idle loop works now (and I'm not sure if you can really change it) is that when you get into it, you're idle no matter what and you simply need to choose an idle state for the CPU to go into. Some code needs to select that state, regardless of what name you want to give to that code. In the current setup, which I really don't think is unreasonable, this is done by cpuidle_select() that simply invokes the governor's ->select() callback and that's it. That callback may very well be part of the scheduler and registered from there if you want that, but why do you want to change the whole mechanism? What's wrong with it now? Further, if you look at your sched_idle(), it looks almost like cpuidle_idle_call() with a few really minor differences (apart from the fact that it doesn't cover suspend-to-idle which it will have to do eventually) that really look arbitrary and the "selection" if () in it simply plays the role of the invocation of ->select(). So how is it different really? > Here, it is the 'predictor' and the API to enter an idle state conforming the idle duration > and the latency constraint. Isn't that just a simple rearrangement of the code? The latency still comes from PM QoS and the duration is computed by your new code instead of that being done by ->select() itself, but why actually ->select() cannot call sched_idle_next_wakeup() to get the duration value it needs? Why do those values need to be passed to a cpuidle_idle_call() replacement as arguments? Is there any particular technical reason for doing that? And why that name, sched_idle_next_wakeup()? Does that function really have anything to do with the scheduler now? > Concerning the testing, it is quite easy to switch from idle_sched to 'menu' > via on sched_debug or whatever option we want to add. > >> >> So I'm still unsure why you want to replace cpuidle_idle_call() with >> sched_idle(). Is there anything wrong with it that it needs to be >> replaced? > > > I don't want to replace cpuidle_idle_call() with sched_idle(). How we > integrate the API is something I would like to discuss with another patchset > focused in this integration only. > > For reference: https://lkml.org/lkml/2016/1/12/530 Please answer my questions above. If you need to post a patchset for this purpose, please do that. I have to say that I was looking forward to the IRQ timings based duration prediction, but the way you want to use it now is seriously disappointing. Thanks, Rafael -- To unsubscribe from this list: send the line "unsubscribe linux-pm" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
Hi Daniel, On 02/16/2016 09:13 PM, Daniel Lezcano wrote: > + > +/** > + * sched_idle_next_wakeup - Predict the next wakeup on the current cpu > + * > + * The next event on the cpu is based on a statistic approach of the > + * interrupt events and the timer deterministic value. From the timer > + * or the irqs, we return the one expected to occur first. > + * > + * Returns the expected remaining idle time before being woken up by > + * an interruption. > + */ > +s64 sched_idle_next_wakeup(void) > +{ > + s64 next_timer = ktime_to_us(tick_nohz_get_sleep_length()); > + s64 next_irq = next_irq_event(); Since next_irq_event() uses RCU and we are in idle this should probably be wrapped in RCU_NONIDLE(). > + > + return min(next_irq, next_timer); > +} > + Thanks, Shreyas -- To unsubscribe from this list: send the line "unsubscribe linux-pm" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On 02/22/2016 04:02 PM, Shreyas B Prabhu wrote: > Hi Daniel, > > On 02/16/2016 09:13 PM, Daniel Lezcano wrote: >> + >> +/** >> + * sched_idle_next_wakeup - Predict the next wakeup on the current cpu >> + * >> + * The next event on the cpu is based on a statistic approach of the >> + * interrupt events and the timer deterministic value. From the timer >> + * or the irqs, we return the one expected to occur first. >> + * >> + * Returns the expected remaining idle time before being woken up by >> + * an interruption. >> + */ >> +s64 sched_idle_next_wakeup(void) >> +{ >> + s64 next_timer = ktime_to_us(tick_nohz_get_sleep_length()); >> + s64 next_irq = next_irq_event(); > > Since next_irq_event() uses RCU and we are in idle this should probably > be wrapped in RCU_NONIDLE(). That is a good point but the function is not supposed to be called in the rcu_idle_enter/rcu_idle_exit section which is inside cpuidle_idle_call().
Hi Rafael, Added Ingo. On 02/20/2016 12:43 AM, Rafael J. Wysocki wrote: > On Fri, Feb 19, 2016 at 4:01 PM, Daniel Lezcano > <daniel.lezcano@linaro.org> wrote: >> On 02/18/2016 07:57 PM, Rafael J. Wysocki wrote: >>> >>> On Thu, Feb 18, 2016 at 11:25 AM, Daniel Lezcano >>> <daniel.lezcano@linaro.org> wrote: >>>> >>>> On 02/17/2016 11:21 PM, Rafael J. Wysocki wrote: >>>> >>>> [ ... ] >>>> >>>>>>> Reviewed-by: Nicolas Pitre <nico@linaro.org> >>>>>> >>>>>> >>>>>> >>>>>> Well, I'm likely overlooking something, but how is this going to be >>>>>> hooked up to the code in idle.c? >>>>> >>>>> >>>>> >>>>> My somewhat educated guess is that sched_idle() in your patch is >>>>> intended to replace cpuidle_idle_call(), right? >>>> >>>> >>>> >>>> Well, no. I was planning to first have it to use a different code path as >>>> experimental code in order to focus improving the accuracy of the >>>> prediction >>>> and then merge or replace cpuidle_idle_call() with sched_idle(). >>> >>> >>> In that case, what about making it a proper cpuidle governor that >>> people can test and play with in a usual way? Then it may potentially >>> benefit everybody and not just your experimental setup and you may get >>> coverage on systems you have no access to normally. >>> >>> There is some boilerplate code to add for this purpose, but that's not >>> that bad IMO. >> >> >> Hi Rafael, >> >> sorry for the delay in the responses. >> >> Actually, adding a new governor is precisely what I would like to avoid >> because the objective is the scheduler acts as the governor. > > But why, really? > > Well, first of all I'm not sure what "the scheduler acts as the > governor" means. For the lack of a better explanation I'll refer to > the message at https://lkml.org/lkml/2016/1/12/530 that you pointed me > at. For better explanation let's refer to Ingo's email: http://lwn.net/Articles/552889/ Ingo said: "Note that I still disagree with the whole design notion of having an "idle back-end" (and a 'cpufreq back end') separate from scheduler power saving policy, and none of the patch-sets offered so far solve this fundamental design problem. [...]" "[...] so the scheduler is in an _ideal_ position to do a judgement call about the near future and estimate how deep an idle state a CPU core should enter into and what frequency it should run at." The patches I sent are supposed to provide the base brick to solve 'this fundamental design problem' by giving a couple of functions: - when is the next event ? - sleep X us with Y latency. The component responsible of choosing X and Y will be the scheduler based on its knowledge of 'the near future' coming from 'when is the next event ?' Moving the code I submitted into a governor has little sense with the goal described above because the sched_next_event will be called from the scheduler code directly. Today the integration is not obvious. There is a branch in the cpuidle loop below but this is just for testing purpose at the moment and not part as a submitted patch. > Your code in there does something like: > > if (sched_idle_enabled()) { > int latency = pm_qos_request(PM_QOS_CPU_DMA_LATENCY); > s64 duration = sched_idle_next_wakeup(); > sched_idle(duration, latency); > } else { > cpuidle_idle_call(); > } > > which is quite questionable to be honest as it adds an extra branch to > the idle loop for no real benefit. That is not the part of the submitted patchset. It is an example based on an embryonic code to show how to use the API. How those API are integrated with the scheduler will be submitted in a separate patchset. > Now, what really is the difference between "governor" and "predictor"? > I don't quite see it except that the former is expected to provide a > specific interface. The predictor is a standalone component giving the answer to the question 'when is the next event ?'. The governor uses the next event information from the prediction to take a decision (governor == scheduler here). If the code I am proposing is buried in the cpuidle framework inside a governor I don't see how the scheduler can become the ideal place to do a decision. > The way the idle loop works now (and I'm not sure if you can really > change it) is that when you get into it, you're idle no matter what > and you simply need to choose an idle state for the CPU to go into. > Some code needs to select that state, regardless of what name you want > to give to that code. > > In the current setup, which I really don't think is unreasonable, this > is done by cpuidle_select() that simply invokes the governor's > ->select() callback and that's it. That callback may very well be > part of the scheduler and registered from there if you want that, but > why do you want to change the whole mechanism? What's wrong with it > now? There is nothing wrong. We lived with it for years and that satisfied everyone. My point is I am sticking to how Ingo and Peter envision the integration work. It is clearly stated in Ingo's mail, he disagree on having the notion of 'idle backend'. The couple of patches here are to be used by the scheduler directly. > Further, if you look at your sched_idle(), it looks almost like > cpuidle_idle_call() with a few really minor differences (apart from > the fact that it doesn't cover suspend-to-idle which it will have to > do eventually) that really look arbitrary and the "selection" if () in > it simply plays the role of the invocation of ->select(). So how is > it different really? The question is not if there is duplicate code or not. Factoring code is easy but factoring code with a moving target is much more difficult. The moving target is the cpuidle code itself. I have been consolidating the code several times and the effort got ruined several times by the changes. It is much more easier to introduce the experimental code make it evolve wisely and then consolidate with the existing code. And it does even more sense to branch it if the feature is by design the exact opposite of the current design (idle based statistics vs irq events) even if there is common code. >> Here, it is the 'predictor' and the API to enter an idle state conforming the idle duration >> and the latency constraint. > > Isn't that just a simple rearrangement of the code? The latency still > comes from PM QoS and the duration is computed by your new code > instead of that being done by ->select() itself, but why actually > ->select() cannot call sched_idle_next_wakeup() to get the duration > value it needs? Why do those values need to be passed to a > cpuidle_idle_call() replacement as arguments? Is there any particular > technical reason for doing that? Yes. The select returns an idle state index. Even if, at some points, we may need to know what is the idle state fitting a set of timing constraints, the scheduler is interested in how long the cpu will be idle in order to take a decision. One example when everything is integrated: The topology gives the information CPU2 and CPU3 are cluster1. 1. CPU3 enters idle, the next wakeup event is stored for this cpu 2. CPU2 enters idle and looks when it should wake up 3. CPU2 looks when CPU3 wakes up 4. CPU2 takes the time intersection between CPU2/CPU3 in order to have the time the cluster is idle 5. CPU2 takes the decision if cluster power down is possible With the current design it is not possible to do that. > And why that name, sched_idle_next_wakeup()? Does that function > really have anything to do with the scheduler now? > >> Concerning the testing, it is quite easy to switch from idle_sched to 'menu' >> via on sched_debug or whatever option we want to add. >> >>> >>> So I'm still unsure why you want to replace cpuidle_idle_call() with >>> sched_idle(). Is there anything wrong with it that it needs to be >>> replaced? >> >> >> I don't want to replace cpuidle_idle_call() with sched_idle(). How we >> integrate the API is something I would like to discuss with another patchset >> focused in this integration only. >> >> For reference: https://lkml.org/lkml/2016/1/12/530 > > Please answer my questions above. If you need to post a patchset for > this purpose, please do that. ... > I have to say that I was looking forward to the IRQ timings based > duration prediction, but the way you want to use it now is seriously > disappointing. This comment is a bit patronizing but I can understand you were expecting more. I will let Peter and Ingo give their opinions and I will follow the consensus. Thanks for the review. -- Daniel
On 02/16/2016 09:13 PM, Daniel Lezcano wrote: [...] > + if (index < 0) { > + /* > + * No idle callbacks fulfilled the constraints, jump > + * to the default function like there wasn't any > + * cpuidle driver. > + */ > + goto default_idle; > + } else { > + /* > + * Enter the idle state previously returned by the > + * governor decision. This function will block until > + * an interrupt occurs and will take care of > + * re-enabling the local interrupts > + */ > + return cpuidle_enter(drv, dev, index); Minor point. You are not calling rcu_idle_exit() in else block. This should probably be ret = cpuidle_enter(drv, dev, index); goto out; > + } > + > +default_idle: > + default_idle_call(); > +out: > + rcu_idle_exit(); > + return ret; > +} > Thanks, Shreyas -- To unsubscribe from this list: send the line "unsubscribe linux-pm" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On 02/22/2016 10:59 PM, Daniel Lezcano wrote: > On 02/22/2016 04:02 PM, Shreyas B Prabhu wrote: >> Hi Daniel, >> >> On 02/16/2016 09:13 PM, Daniel Lezcano wrote: >>> + >>> +/** >>> + * sched_idle_next_wakeup - Predict the next wakeup on the current cpu >>> + * >>> + * The next event on the cpu is based on a statistic approach of the >>> + * interrupt events and the timer deterministic value. From the timer >>> + * or the irqs, we return the one expected to occur first. >>> + * >>> + * Returns the expected remaining idle time before being woken up by >>> + * an interruption. >>> + */ >>> +s64 sched_idle_next_wakeup(void) >>> +{ >>> + s64 next_timer = ktime_to_us(tick_nohz_get_sleep_length()); >>> + s64 next_irq = next_irq_event(); >> >> Since next_irq_event() uses RCU and we are in idle this should probably >> be wrapped in RCU_NONIDLE(). > > That is a good point but the function is not supposed to be called in > the rcu_idle_enter/rcu_idle_exit section which is inside > cpuidle_idle_call(). > > Right. My bad. -- To unsubscribe from this list: send the line "unsubscribe linux-pm" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On 02/23/2016 11:06 AM, Shreyas B Prabhu wrote: > > > On 02/16/2016 09:13 PM, Daniel Lezcano wrote: > [...] >> + if (index < 0) { >> + /* >> + * No idle callbacks fulfilled the constraints, jump >> + * to the default function like there wasn't any >> + * cpuidle driver. >> + */ >> + goto default_idle; >> + } else { >> + /* >> + * Enter the idle state previously returned by the >> + * governor decision. This function will block until >> + * an interrupt occurs and will take care of >> + * re-enabling the local interrupts >> + */ >> + return cpuidle_enter(drv, dev, index); > > Minor point. You are not calling rcu_idle_exit() in else block. This > should probably be > ret = cpuidle_enter(drv, dev, index); > goto out; Yes. Right. Thanks for the review. -- Daniel
diff --git a/drivers/cpuidle/Kconfig b/drivers/cpuidle/Kconfig index 7e48eb5..d64445b 100644 --- a/drivers/cpuidle/Kconfig +++ b/drivers/cpuidle/Kconfig @@ -23,6 +23,15 @@ config CPU_IDLE_GOV_LADDER config CPU_IDLE_GOV_MENU bool "Menu governor (for tickless system)" +config CPU_IDLE_GOV_SCHED + bool "Sched idle governor" + select IRQ_TIMINGS + help + Enables an irq timings tracking mechanism to track the wakeup sources + of the platform. + + If you are unsure, it is safe to say N. + config DT_IDLE_STATES bool diff --git a/kernel/sched/Makefile b/kernel/sched/Makefile index 6768797..f7d5a35 100644 --- a/kernel/sched/Makefile +++ b/kernel/sched/Makefile @@ -19,3 +19,4 @@ obj-$(CONFIG_SCHED_AUTOGROUP) += auto_group.o obj-$(CONFIG_SCHEDSTATS) += stats.o obj-$(CONFIG_SCHED_DEBUG) += debug.o obj-$(CONFIG_CGROUP_CPUACCT) += cpuacct.o +obj-$(CONFIG_CPU_IDLE_GOV_SCHED) += idle-sched.o diff --git a/kernel/sched/idle-sched.c b/kernel/sched/idle-sched.c new file mode 100644 index 0000000..4dc03da --- /dev/null +++ b/kernel/sched/idle-sched.c @@ -0,0 +1,280 @@ +/* + * Copyright (C) 2016 Linaro Ltd, Daniel Lezcano <daniel.lezcano@linaro.org> + * Nicolas Pitre <nicolas.pitre@linaro.org> + * + * This program is free software; you can redistribute it and/or modify + * it under the terms of the GNU General Public License version 2 as + * published by the Free Software Foundation. + * + */ +#include <linux/cpuidle.h> +#include <linux/interrupt.h> +#include <linux/ktime.h> +#include <linux/tick.h> + +/** + * irqt_mean - compute the average + * + * @irqt: the irq timings structure + * + * Returns an u32 corresponding to the mean value, or zero if there is + * no data + */ +static inline u32 irqt_mean(struct irq_timings *irqt) +{ + return irqt->sum >> IRQ_TIMINGS_SHIFT; +} + +/** + * irqt_variance - compute the variance + * + * @irqt: the irq timings structure + * + * Returns an u64 corresponding to the variance, or zero if there is + * no data + */ +static u64 irqt_variance(struct irq_timings *irqt, u32 mean) +{ + int i; + u64 variance = 0; + + /* + * The variance is the sum of the squared difference to the + * average divided by the number of elements. + */ + for (i = 0; i < IRQ_TIMINGS_SIZE; i++) { + s32 diff = irqt->values[i] - mean; + variance += (s64)diff * diff; + } + + return variance >> IRQ_TIMINGS_SHIFT; +} + +/** + * next_irq_event - returns the next irq event + * + * Returns a guess estimate remaining time before an interrupt + * occurs. The type s64 corresponding in a duration in microsecond. + * + * The value is always greater or equal to zero. In case, there is no + * possible prediction, S64_MAX is returned. + */ +static s64 next_irq_event(void) +{ + struct irq_timings *irqt; + unsigned int irq = 0; + u64 variance, now; + s64 diff, next = 0, min = S64_MAX; + u32 interval, mean; + s32 deviation; + + now = local_clock(); + + /* + * Search for the earliest expected interruption. + */ + rcu_read_lock(); + while ((irqt = irqtiming_get_next(&irq))) { + + /* + * No values available. + */ + if (!irqt->timestamp) + continue; + + /* + * This interrupt last triggered more than a second ago. + * It is definitely not predictable for our purpose anymore. + */ + if ((now - irqt->timestamp) > NSEC_PER_SEC) + continue; + + /* + * If the mean value is null, just ignore this wakeup + * source. + */ + mean = irqt_mean(irqt); + if (!mean) + continue; + + variance = irqt_variance(irqt, mean); + /* + * We want to check the last interval is: + * + * mean - stddev < interval < mean + stddev + * + * That simplifies to: + * + * -stddev < interval - mean < stddev + * + * abs(interval - mean) < stddev + * + * The standard deviation is the sqrt of the variance: + * + * abs(interval - mean) < sqrt(variance) + * + * and we want to avoid the sqrt, so we square the + * equation: + * + * (interval - mean)^2 < variance + * + * So if the latest value of the stats complies with + * this condition, then the wakeup source is + * considered predictable and can be used to predict + * the next event. + */ + interval = irqt->values[irqt->w_index]; + deviation = interval - mean; + if ((s64)deviation * deviation > variance) + continue; + + /* + * Let's compute the next event: the wakeup source is + * considered predictable, we add the average interval + * time added to the latest interruption event + * time. Note the timestamp is stored in nsec while + * the mean time is microsec. + */ + next = irqt->timestamp + (mean << 10); + + /* + * If the interrupt is supposed to happen before the + * minimum time, then it becomes the minimum. + */ + if (next < min) + min = next; + } + rcu_read_unlock(); + + /* + * There is no prediction for any interrupt, so we return + * S64_MAX. + */ + if (!next) + return S64_MAX; + + /* + * At this point, we have our prediction but the caller is + * expecting the remaining time before the next event, so + * compute the expected sleep length. + */ + diff = min - now; + + /* + * min and now are nanosecond units, let's convert to + * microsec the result of the diff. + */ + diff >>= 10; + + /* + * The result could be negative for different reasons: + * - the prediction is incorrect but we can't do anything here + * except assuming the interrupt occured effectively. + * + * - the prediction was too near now and expired while we were + * in this function. + * + * In those cases, we return 0. Otherwise we return the expected + * duration before an interrupt occurs. + */ + return diff > 0 ? diff : 0; +} + +/** + * sched_idle_next_wakeup - Predict the next wakeup on the current cpu + * + * The next event on the cpu is based on a statistic approach of the + * interrupt events and the timer deterministic value. From the timer + * or the irqs, we return the one expected to occur first. + * + * Returns the expected remaining idle time before being woken up by + * an interruption. + */ +s64 sched_idle_next_wakeup(void) +{ + s64 next_timer = ktime_to_us(tick_nohz_get_sleep_length()); + s64 next_irq = next_irq_event(); + + return min(next_irq, next_timer); +} + +/** + * sched_idle - go to idle for a specified amount of time + * + * @duration: the idle duration time + * @latency: the latency constraint + * + * Returns 0 on success, < 0 otherwise. + */ +int sched_idle(s64 duration, unsigned int latency) +{ + struct cpuidle_device *dev = __this_cpu_read(cpuidle_devices); + struct cpuidle_driver *drv = cpuidle_get_cpu_driver(dev); + struct cpuidle_state_usage *su; + struct cpuidle_state *s; + int i, ret = 0, index = -1; + + rcu_idle_enter(); + + /* + * No cpuidle driver is available, let's use the default arch + * idle function. + */ + if (cpuidle_not_available(drv, dev)) + goto default_idle; + + /* + * Find the idle state with the lowest power while satisfying + * our constraints. We will save energy if the duration of the + * idle time is bigger than the target residency which is the + * break even point. The choice will be modulated by the + * latency. + */ + for (i = 0; i < drv->state_count; i++) { + + s = &drv->states[i]; + + su = &dev->states_usage[i]; + + if (s->disabled || su->disable) + continue; + if (s->target_residency > duration) + continue; + if (s->exit_latency > latency) + continue; + + index = i; + } + + /* + * The idle task must be scheduled, it is pointless to go to + * idle, just re-enable the interrupt and return. + */ + if (current_clr_polling_and_test()) { + local_irq_enable(); + goto out; + } + + if (index < 0) { + /* + * No idle callbacks fulfilled the constraints, jump + * to the default function like there wasn't any + * cpuidle driver. + */ + goto default_idle; + } else { + /* + * Enter the idle state previously returned by the + * governor decision. This function will block until + * an interrupt occurs and will take care of + * re-enabling the local interrupts + */ + return cpuidle_enter(drv, dev, index); + } + +default_idle: + default_idle_call(); +out: + rcu_idle_exit(); + return ret; +}
Many IRQs are quiet most of the time, or they tend to come in bursts of fairly equal time intervals within each burst. It is therefore possible to detect those IRQs with stable intervals and guestimate when the next IRQ event is most likely to happen. Examples of such IRQs may include audio related IRQs where the FIFO size and/or DMA descriptor size with the sample rate create stable intervals, block devices during large data transfers, etc. Even network streaming of multimedia content creates patterns of periodic network interface IRQs in some cases. This patch adds code to compute the mean interval and variance for each IRQ over a window of time intervals between IRQ events. Those statistics can be used to assist cpuidle in selecting the most appropriate sleep state by predicting the most likely time for the next interrupt. Signed-off-by: Daniel Lezcano <daniel.lezcano@linaro.org> --- drivers/cpuidle/Kconfig | 9 ++ kernel/sched/Makefile | 1 + kernel/sched/idle-sched.c | 280 ++++++++++++++++++++++++++++++++++++++++++++++ 3 files changed, 290 insertions(+) create mode 100644 kernel/sched/idle-sched.c