| Message ID | 1499189651-18797-1-git-send-email-patrick.bellasi@arm.com |
|---|---|
| Headers | show |
| Series | cpufreq: schedutil: fixes for flags updates | expand |
On 04-07-17, 18:34, Patrick Bellasi wrote: > diff --git a/include/linux/sched/cpufreq.h b/include/linux/sched/cpufreq.h > index d2be2cc..36ac8d2 100644 > --- a/include/linux/sched/cpufreq.h > +++ b/include/linux/sched/cpufreq.h > @@ -10,6 +10,7 @@ > #define SCHED_CPUFREQ_RT (1U << 0) > #define SCHED_CPUFREQ_DL (1U << 1) > #define SCHED_CPUFREQ_IOWAIT (1U << 2) > +#define SCHED_CPUFREQ_IDLE (1U << 3) > > #define SCHED_CPUFREQ_RT_DL (SCHED_CPUFREQ_RT | SCHED_CPUFREQ_DL) > > diff --git a/kernel/sched/cpufreq_schedutil.c b/kernel/sched/cpufreq_schedutil.c > index eaba6d6..004ae18 100644 > --- a/kernel/sched/cpufreq_schedutil.c > +++ b/kernel/sched/cpufreq_schedutil.c > @@ -304,6 +304,12 @@ static void sugov_update_shared(struct update_util_data *hook, u64 time, > > sg_cpu->util = util; > sg_cpu->max = max; > + > + /* CPU is entering IDLE, reset flags without triggering an update */ > + if (unlikely(flags & SCHED_CPUFREQ_IDLE)) { > + sg_cpu->flags = 0; > + goto done; > + } Why is it important to have the above diff at all ? For example we aren't doing similar stuff in sugov_update_single() and that will go on and try to change the frequency if rate_limit_us time is over since last update. And also why is it important to write 0 to sg_cpu->flags ? What wouldn't work if we set sg_cpu->flags to SCHED_CPUFREQ_IDLE in this case ? i.e. Just the below statement should be good for us. > sg_cpu->flags = flags; > > sugov_set_iowait_boost(sg_cpu, time, flags); > @@ -318,6 +324,7 @@ static void sugov_update_shared(struct update_util_data *hook, u64 time, > sugov_update_commit(sg_policy, time, next_f); > } > > +done: > raw_spin_unlock(&sg_policy->update_lock); > } > > diff --git a/kernel/sched/idle_task.c b/kernel/sched/idle_task.c > index 0c00172..a844c91 100644 > --- a/kernel/sched/idle_task.c > +++ b/kernel/sched/idle_task.c > @@ -29,6 +29,10 @@ pick_next_task_idle(struct rq *rq, struct task_struct *prev, struct rq_flags *rf > put_prev_task(rq, prev); > update_idle_core(rq); > schedstat_inc(rq->sched_goidle); > + > + /* kick cpufreq (see the comment in kernel/sched/sched.h). */ > + cpufreq_update_this_cpu(rq, SCHED_CPUFREQ_IDLE); > + This looks correct. Can we completely avoid the utilization contribution of the CPUs which have gone idle? Right now we avoid them with help of (delta_ns > TICK_NSEC). Can we instead check this SCHED_CPUFREQ_IDLE flag ? -- viresh
On 04-07-17, 18:34, Patrick Bellasi wrote: > In system where multiple CPUs shares the same frequency domain a small > workload on a CPU can still be subject to frequency spikes, generated by > the activation of the sugov's kthread. > > Since the sugov kthread is a special RT task, which goal is just that to > activate a frequency transition, it does not make sense for it to bias > the schedutil's frequency selection policy. > > This patch exploits the information related to the current task to silently > ignore cpufreq_update_this_cpu() calls, coming from the RT scheduler, while > the sugov kthread is running. > > Signed-off-by: Patrick Bellasi <patrick.bellasi@arm.com> > Cc: Ingo Molnar <mingo@redhat.com> > Cc: Peter Zijlstra <peterz@infradead.org> > Cc: Rafael J. Wysocki <rafael.j.wysocki@intel.com> > Cc: Viresh Kumar <viresh.kumar@linaro.org> > Cc: linux-kernel@vger.kernel.org > Cc: linux-pm@vger.kernel.org > > --- > Changes from v1: > - move check before policy spinlock (JuriL) > --- > kernel/sched/cpufreq_schedutil.c | 8 ++++++++ > 1 file changed, 8 insertions(+) > > diff --git a/kernel/sched/cpufreq_schedutil.c b/kernel/sched/cpufreq_schedutil.c > index c982dd0..eaba6d6 100644 > --- a/kernel/sched/cpufreq_schedutil.c > +++ b/kernel/sched/cpufreq_schedutil.c > @@ -218,6 +218,10 @@ static void sugov_update_single(struct update_util_data *hook, u64 time, > unsigned int next_f; > bool busy; > > + /* Skip updates generated by sugov kthreads */ > + if (unlikely(current == sg_policy->thread)) > + return; > + > sugov_set_iowait_boost(sg_cpu, time, flags); > sg_cpu->last_update = time; > > @@ -290,6 +294,10 @@ static void sugov_update_shared(struct update_util_data *hook, u64 time, > unsigned long util, max; > unsigned int next_f; > > + /* Skip updates generated by sugov kthreads */ > + if (unlikely(current == sg_policy->thread)) > + return; > + > sugov_get_util(&util, &max); Yes we discussed this last time as well (I looked again at those discussions and am still confused a bit), but wanted to clarify one more time. After the 2nd patch of this series is applied, why will we still have this problem? As we concluded it last time, the problem wouldn't happen until the time the sugov RT thread is running (Hint: work_in_progress). And once the sugov RT thread is gone, one of the other scheduling classes will take over and should update the flag pretty quickly. Are we worried about the time between the sugov RT thread finishes and when the CFS or IDLE sched class call the util handler again? If yes, then we will still have that problem for any normal RT/DL task. Isn't it ? -- viresh
On 05-Jul 10:30, Viresh Kumar wrote: > On 04-07-17, 18:34, Patrick Bellasi wrote: > > In system where multiple CPUs shares the same frequency domain a small > > workload on a CPU can still be subject to frequency spikes, generated by > > the activation of the sugov's kthread. > > > > Since the sugov kthread is a special RT task, which goal is just that to > > activate a frequency transition, it does not make sense for it to bias > > the schedutil's frequency selection policy. > > > > This patch exploits the information related to the current task to silently > > ignore cpufreq_update_this_cpu() calls, coming from the RT scheduler, while > > the sugov kthread is running. > > > > Signed-off-by: Patrick Bellasi <patrick.bellasi@arm.com> > > Cc: Ingo Molnar <mingo@redhat.com> > > Cc: Peter Zijlstra <peterz@infradead.org> > > Cc: Rafael J. Wysocki <rafael.j.wysocki@intel.com> > > Cc: Viresh Kumar <viresh.kumar@linaro.org> > > Cc: linux-kernel@vger.kernel.org > > Cc: linux-pm@vger.kernel.org > > > > --- > > Changes from v1: > > - move check before policy spinlock (JuriL) > > --- > > kernel/sched/cpufreq_schedutil.c | 8 ++++++++ > > 1 file changed, 8 insertions(+) > > > > diff --git a/kernel/sched/cpufreq_schedutil.c b/kernel/sched/cpufreq_schedutil.c > > index c982dd0..eaba6d6 100644 > > --- a/kernel/sched/cpufreq_schedutil.c > > +++ b/kernel/sched/cpufreq_schedutil.c > > @@ -218,6 +218,10 @@ static void sugov_update_single(struct update_util_data *hook, u64 time, > > unsigned int next_f; > > bool busy; > > > > + /* Skip updates generated by sugov kthreads */ > > + if (unlikely(current == sg_policy->thread)) > > + return; > > + > > sugov_set_iowait_boost(sg_cpu, time, flags); > > sg_cpu->last_update = time; > > > > @@ -290,6 +294,10 @@ static void sugov_update_shared(struct update_util_data *hook, u64 time, > > unsigned long util, max; > > unsigned int next_f; > > > > + /* Skip updates generated by sugov kthreads */ > > + if (unlikely(current == sg_policy->thread)) > > + return; > > + > > sugov_get_util(&util, &max); > > Yes we discussed this last time as well (I looked again at those discussions and > am still confused a bit), but wanted to clarify one more time. > > After the 2nd patch of this series is applied, why will we still have this > problem? As we concluded it last time, the problem wouldn't happen until the > time the sugov RT thread is running (Hint: work_in_progress). And once the sugov > RT thread is gone, one of the other scheduling classes will take over and should > update the flag pretty quickly. > > Are we worried about the time between the sugov RT thread finishes and when the > CFS or IDLE sched class call the util handler again? If yes, then we will still > have that problem for any normal RT/DL task. Isn't it ? Yes, we are worried about that time, without this we can generate spikes to the max OPP even when only relatively small FAIR tasks are running. The same problem is not there for the other "normal RT/DL" tasks, just because for those tasks this is the expected behavior: we wanna go to max. To the contrary the sugov kthread, although being a RT task, is just functional to the "machinery" to work, it's an actuator. Thus, IMO it makes no sense from a design standpoint for it to interfere whatsoever with what the "machinery" is doing. Finally, the second patch of this series fixes a kind-of symmetrical issue: while this one avoid going to max OPP, the next one avoid to stay at max OPP once not more needed. Cheers Patrick -- #include <best/regards.h> Patrick Bellasi
On 05-Jul 10:20, Viresh Kumar wrote: > On 04-07-17, 18:34, Patrick Bellasi wrote: > > diff --git a/include/linux/sched/cpufreq.h b/include/linux/sched/cpufreq.h > > index d2be2cc..36ac8d2 100644 > > --- a/include/linux/sched/cpufreq.h > > +++ b/include/linux/sched/cpufreq.h > > @@ -10,6 +10,7 @@ > > #define SCHED_CPUFREQ_RT (1U << 0) > > #define SCHED_CPUFREQ_DL (1U << 1) > > #define SCHED_CPUFREQ_IOWAIT (1U << 2) > > +#define SCHED_CPUFREQ_IDLE (1U << 3) > > > > #define SCHED_CPUFREQ_RT_DL (SCHED_CPUFREQ_RT | SCHED_CPUFREQ_DL) > > > > diff --git a/kernel/sched/cpufreq_schedutil.c b/kernel/sched/cpufreq_schedutil.c > > index eaba6d6..004ae18 100644 > > --- a/kernel/sched/cpufreq_schedutil.c > > +++ b/kernel/sched/cpufreq_schedutil.c > > @@ -304,6 +304,12 @@ static void sugov_update_shared(struct update_util_data *hook, u64 time, > > > > sg_cpu->util = util; > > sg_cpu->max = max; > > + > > + /* CPU is entering IDLE, reset flags without triggering an update */ > > + if (unlikely(flags & SCHED_CPUFREQ_IDLE)) { > > + sg_cpu->flags = 0; > > + goto done; > > + } > > Why is it important to have the above diff at all ? For example we aren't doing > similar stuff in sugov_update_single() and that will go on and try to change the > frequency if rate_limit_us time is over since last update. The p repose here is mainly to avoid interference of IDLE CPUs on other CPUs in the same frequency domain, by just resetting their "requests". In the single case, it's completely up to the policy to decide what to do when we enter IDLE without risking to affect other CPUs. But perhaps you are right, maybe we should use the same heuristics in both cases. Entering idle just reset the flags and do not enforce for example a frequency drop. > And also why is it important to write 0 to sg_cpu->flags ? What wouldn't work if > we set sg_cpu->flags to SCHED_CPUFREQ_IDLE in this case ? i.e. Just the below > statement should be good for us. Let say flags have the RT/DL flag set when the RT task sleep, is there any specific reason to keep this flag up while the CPU is IDLE? IOW, why should we care about an information related to an even which is now over? The proposal of this patch is just meant to make sure that the flags, being a state variable, always describe the current status of the sugov "state machine". If a CPU is IDLE there are not sensible events going on and thus flags should better be reset. > > > sg_cpu->flags = flags; > > > > sugov_set_iowait_boost(sg_cpu, time, flags); > > @@ -318,6 +324,7 @@ static void sugov_update_shared(struct update_util_data *hook, u64 time, > > sugov_update_commit(sg_policy, time, next_f); > > } > > > > +done: > > raw_spin_unlock(&sg_policy->update_lock); > > } > > > > diff --git a/kernel/sched/idle_task.c b/kernel/sched/idle_task.c > > index 0c00172..a844c91 100644 > > --- a/kernel/sched/idle_task.c > > +++ b/kernel/sched/idle_task.c > > @@ -29,6 +29,10 @@ pick_next_task_idle(struct rq *rq, struct task_struct *prev, struct rq_flags *rf > > put_prev_task(rq, prev); > > update_idle_core(rq); > > schedstat_inc(rq->sched_goidle); > > + > > + /* kick cpufreq (see the comment in kernel/sched/sched.h). */ > > + cpufreq_update_this_cpu(rq, SCHED_CPUFREQ_IDLE); > > + > > This looks correct. > > Can we completely avoid the utilization contribution of the CPUs which have gone > idle? Right now we avoid them with help of (delta_ns > TICK_NSEC). Can we > instead check this SCHED_CPUFREQ_IDLE flag ? I would say that the blocked utilization of an IDLE CPU is still worth to be considered, at least for a limited amount of time, for few main reasons: 1. it represents CPU bandwidth that is likely to be required by a task which can wakeup in a short while. Consider for example an 80% task activated every 16ms: even if it's not running right now it's likely to wakeup in the next ~3ms to run for the following ~13ms. Thus, we should probably better consider that CPU utilization. 2. we already have policies to gratefully reduce the current OPP if its utilization decrease. This means that we are interested in a sort of policy which favors higher OPPs to avoid impacting performance of tasks which suddenly wakeup. 3. A CPU entering IDLE is not a great source of new information for OPP selection, I would not strictly bind an OPP change to this event. That's also why this patch propose to clear the flags without actually triggering an OPP change. Moreover, maybe the issue you are trying to solve it's more related to having a stale utilization for an IDLE CPUs? In that case we should fix the real source of the issue, which is the utilization of an IDLE CPU not being updated over time. But that's outside of the scope of this series. Cheers Patrick -- #include <best/regards.h> Patrick Bellasi
[+Brendan] On 05-Jul 11:31, Viresh Kumar wrote: > On 04-07-17, 18:34, Patrick Bellasi wrote: > > diff --git a/kernel/sched/cpufreq_schedutil.c b/kernel/sched/cpufreq_schedutil.c > > index 004ae18..98704d8 100644 > > --- a/kernel/sched/cpufreq_schedutil.c > > +++ b/kernel/sched/cpufreq_schedutil.c > > @@ -216,6 +216,7 @@ static void sugov_update_single(struct update_util_data *hook, u64 time, > > struct cpufreq_policy *policy = sg_policy->policy; > > unsigned long util, max; > > unsigned int next_f; > > + bool rt_mode; > > bool busy; > > > > /* Skip updates generated by sugov kthreads */ > > @@ -230,7 +231,15 @@ static void sugov_update_single(struct update_util_data *hook, u64 time, > > > > busy = sugov_cpu_is_busy(sg_cpu); > > > > - if (flags & SCHED_CPUFREQ_RT_DL) { > > + /* > > + * While RT/DL tasks are running we do not want FAIR tasks to > > + * overvrite this CPU's flags, still we can update utilization and > > + * frequency (if required/possible) to be fair with these tasks. > > + */ > > + rt_mode = task_has_dl_policy(current) || > > + task_has_rt_policy(current) || > > + (flags & SCHED_CPUFREQ_RT_DL); > > We may want to create a separate inline function for above, as it is already > used twice in this patch. Good idea. > But I was wondering if we can get some help from the scheduler to avoid such > code here. I understand that we don't want to do the aggregation in the > scheduler to keep it clean and keep such governor specific thing here. Indeed. > But what about clearing the sched-class's flag from .pick_next_task() callback > when they return NULL ? > > What about something like this instead (completely untested), with which we > don't need the 2/3 patch as well: Just had a fast check but I think something like that can work. We had an internal discussion with Brendan (in CC now) which had a similar proposal. Main counter arguments for me was: 1. we wanna to reduce the pollution of scheduling classes code with schedutil specific code, unless strictly necessary 2. we never had a "clear bit" semantics for flags updates Thus this proposal seemed to me less of a discontinuity wrt the current interface. However, something similar to what you propose below should also work. Let's collect some more feedback... > diff --git a/include/linux/sched/cpufreq.h b/include/linux/sched/cpufreq.h > index d2be2ccbb372..e81a6b5591f5 100644 > --- a/include/linux/sched/cpufreq.h > +++ b/include/linux/sched/cpufreq.h > @@ -11,6 +11,10 @@ > #define SCHED_CPUFREQ_DL (1U << 1) > #define SCHED_CPUFREQ_IOWAIT (1U << 2) > > +#define SCHED_CPUFREQ_CLEAR (1U << 31) > +#define SCHED_CPUFREQ_CLEAR_RT (SCHED_CPUFREQ_CLEAR | SCHED_CPUFREQ_RT) > +#define SCHED_CPUFREQ_CLEAR_DL (SCHED_CPUFREQ_CLEAR | SCHED_CPUFREQ_DL) > + > #define SCHED_CPUFREQ_RT_DL (SCHED_CPUFREQ_RT | SCHED_CPUFREQ_DL) > > #ifdef CONFIG_CPU_FREQ > diff --git a/kernel/sched/cpufreq_schedutil.c b/kernel/sched/cpufreq_schedutil.c > index 076a2e31951c..f32e15d59d62 100644 > --- a/kernel/sched/cpufreq_schedutil.c > +++ b/kernel/sched/cpufreq_schedutil.c > @@ -218,6 +218,9 @@ static void sugov_update_single(struct update_util_data *hook, u64 time, > unsigned int next_f; > bool busy; > > + if (flags & SCHED_CPUFREQ_CLEAR) > + return; Here we should still clear the flags, like what we do for the shared case... just to keep the internal status consiste with the notifications we have got from the scheduling classes. > + > sugov_set_iowait_boost(sg_cpu, time, flags); > sg_cpu->last_update = time; > > @@ -296,7 +299,13 @@ static void sugov_update_shared(struct update_util_data *hook, u64 time, > > sg_cpu->util = util; > sg_cpu->max = max; > - sg_cpu->flags = flags; > + > + if (unlikely(flags & SCHED_CPUFREQ_CLEAR)) { > + sg_cpu->flags &= ~(flags & ~SCHED_CPUFREQ_CLEAR); > + return; > + } > + > + sg_cpu->flags |= flags; > > sugov_set_iowait_boost(sg_cpu, time, flags); > sg_cpu->last_update = time; > diff --git a/kernel/sched/deadline.c b/kernel/sched/deadline.c > index a2ce59015642..441d6153d654 100644 > --- a/kernel/sched/deadline.c > +++ b/kernel/sched/deadline.c > @@ -1203,8 +1203,10 @@ pick_next_task_dl(struct rq *rq, struct task_struct *prev, struct rq_flags *rf) > if (prev->sched_class == &dl_sched_class) > update_curr_dl(rq); > > - if (unlikely(!dl_rq->dl_nr_running)) > + if (unlikely(!dl_rq->dl_nr_running)) { > + cpufreq_update_this_cpu(rq, SCHED_CPUFREQ_CLEAR_DL); > return NULL; > + } > > put_prev_task(rq, prev); > > diff --git a/kernel/sched/rt.c b/kernel/sched/rt.c > index 979b7341008a..bca9e4bb7ec4 100644 > --- a/kernel/sched/rt.c > +++ b/kernel/sched/rt.c > @@ -1556,8 +1556,10 @@ pick_next_task_rt(struct rq *rq, struct task_struct *prev, struct rq_flags *rf) > if (prev->sched_class == &rt_sched_class) > update_curr_rt(rq); > > - if (!rt_rq->rt_queued) > + if (!rt_rq->rt_queued) { > + cpufreq_update_this_cpu(rq, SCHED_CPUFREQ_CLEAR_RT); > return NULL; > + } > > put_prev_task(rq, prev); > > -- > viresh -- #include <best/regards.h> Patrick Bellasi
On 05-07-17, 14:04, Patrick Bellasi wrote: > On 05-Jul 10:20, Viresh Kumar wrote: > > And also why is it important to write 0 to sg_cpu->flags ? What wouldn't work if > > we set sg_cpu->flags to SCHED_CPUFREQ_IDLE in this case ? i.e. Just the below > > statement should be good for us. > > Let say flags have the RT/DL flag set when the RT task sleep, is there > any specific reason to keep this flag up while the CPU is IDLE? > IOW, why should we care about an information related to an even which > is now over? Maybe I wasn't able to communicate what I wanted to say, but I am not asking you to keep RT/DL flags as is, but rather set the flags variable to SCHED_CPUFREQ_IDLE (1 << 3). My concerns were about adding an additional conditional statement here, while we can live without one. > The proposal of this patch is just meant to make sure that the flags, > being a state variable, always describe the current status of the > sugov "state machine". > If a CPU is IDLE there are not sensible events going on and thus flags > should better be reset. or set to SCHED_CPUFREQ_IDLE. > > This looks correct. > > > > Can we completely avoid the utilization contribution of the CPUs which have gone > > idle? Right now we avoid them with help of (delta_ns > TICK_NSEC). Can we > > instead check this SCHED_CPUFREQ_IDLE flag ? > > I would say that the blocked utilization of an IDLE CPU is still worth > to be considered, at least for a limited amount of time, for few main > reasons: > > 1. it represents CPU bandwidth that is likely to be required by a task > which can wakeup in a short while. Consider for example an 80% task > activated every 16ms: even if it's not running right now it's > likely to wakeup in the next ~3ms to run for the following ~13ms. > Thus, we should probably better consider that CPU utilization. > > 2. we already have policies to gratefully reduce the current OPP if > its utilization decrease. This means that we are interested in a > sort of policy which favors higher OPPs to avoid impacting > performance of tasks which suddenly wakeup. > > 3. A CPU entering IDLE is not a great source of new information > for OPP selection, I would not strictly bind an OPP change to this > event. That's also why this patch propose to clear the flags > without actually triggering an OPP change. > > Moreover, maybe the issue you are trying to solve it's more related to > having a stale utilization for an IDLE CPUs? I wasn't trying to solve any issue here, but just discussing about what should we do here. Yeah it seems fair to keep the utilization of the idle CPU for another TICK, after which we are ignoring it anyway. -- viresh
On Tuesday, July 04, 2017 06:34:05 PM Patrick Bellasi wrote: > Each time a CPU utilisation update is issued by the scheduler a flag, which > mainly defines which scheduling class is asking for the update, is used by the > frequency selection policy to support the selection of the most appropriate > OPP. > > In the current implementation, CPU flags are overridden each time the scheduler > calls schedutil for an update. Such a behavior seems to be sub-optimal, > especially on systems where frequency domains span across multiple CPUs. > > Indeed, assuming CPU1 and CPU2 share the same frequency domain, there can be > the following issues: > > A) Small FAIR task running at MAX OPP. > A RT task, which just executed on CPU1, can keep the domain at the > max frequency for a prolonged period of time after its completion, > even if there are no longer RT tasks running on CPUs of its domain. > > B) FAIR wakeup reducing the OPP of the current RT task. > A FAIR task enqueued in a CPU where a RT task is running overrides the flag > configured by the RT task thus potentially causing an unwanted frequency > drop. > > C) RT wakeup not running at max OPP. > An RT task waking up on a CPU which has recently updated its OPP can > be forced to run at a lower frequency because of the throttling > enforced by schedutil, even if there are not OPP transitions > currently in progress. > > .:: Patches organization > ======================== > > This series proposes a set of fixes for the aforementioned issues and it's an > update addressing all the main comments collected from the previous posting > [1]. It seems to me that there is a nonzero overlap between this and the Juri's work. If that's correct, I'd like this series to go on top of the Juri's one. Thanks, Rafael
On Tue, Jul 4, 2017 at 10:34 AM, Patrick Bellasi <patrick.bellasi@arm.com> wrote: > Currently, sg_cpu's flags are set to the value defined by the last call of > the cpufreq_update_util()/cpufreq_update_this_cpu(); for RT/DL classes > this corresponds to the SCHED_CPUFREQ_{RT/DL} flags always being set. > > When multiple CPU shares the same frequency domain it might happen that a > CPU which executed a RT task, right before entering IDLE, has one of the > SCHED_CPUFREQ_RT_DL flags set, permanently, until it exits IDLE. > > Although such an idle CPU is _going to be_ ignored by the > sugov_next_freq_shared(): > 1. this kind of "useless RT requests" are ignored only if more then > TICK_NSEC have elapsed since the last update > 2. we can still potentially trigger an already too late switch to > MAX, which starts also a new throttling interval > 3. the internal state machine is not consistent with what the > scheduler knows, i.e. the CPU is now actually idle > > Thus, in sugov_next_freq_shared(), where utilisation and flags are > aggregated across all the CPUs of a frequency domain, it can turn out > that all the CPUs of that domain can run unnecessary at the maximum OPP > until another event happens in the idle CPU, which eventually clear the > SCHED_CPUFREQ_{RT/DL} flag, or the IDLE CPUs gets ignored after > TICK_NSEC since the CPU entering IDLE. > > Such a behaviour can harm the energy efficiency of systems where RT > workloads are not so frequent and other CPUs in the same frequency > domain are running small utilisation workloads, which is a quite common > scenario in mobile embedded systems. > > This patch proposes a solution which is aligned with the current principle > to update the flags each time a scheduling event happens. The scheduling > of the idle_task on a CPU is considered one of such meaningful events. > That's why when the idle_task is selected for execution we poke the > schedutil policy to reset the flags for that CPU. > > No frequency transitions are activated at that point, which is fair in > case the RT workload should come back in the future. However, this still > allows other CPUs in the same frequency domain to scale down the > frequency in case that should be possible. > > Signed-off-by: Patrick Bellasi <patrick.bellasi@arm.com> > Cc: Ingo Molnar <mingo@redhat.com> > Cc: Peter Zijlstra <peterz@infradead.org> > Cc: Rafael J. Wysocki <rafael.j.wysocki@intel.com> > Cc: Viresh Kumar <viresh.kumar@linaro.org> > Cc: linux-kernel@vger.kernel.org > Cc: linux-pm@vger.kernel.org > > --- > Changes from v1: > - added "unlikely()" around the statement (SteveR) > --- > include/linux/sched/cpufreq.h | 1 + > kernel/sched/cpufreq_schedutil.c | 7 +++++++ > kernel/sched/idle_task.c | 4 ++++ > 3 files changed, 12 insertions(+) > > diff --git a/include/linux/sched/cpufreq.h b/include/linux/sched/cpufreq.h > index d2be2cc..36ac8d2 100644 > --- a/include/linux/sched/cpufreq.h > +++ b/include/linux/sched/cpufreq.h > @@ -10,6 +10,7 @@ > #define SCHED_CPUFREQ_RT (1U << 0) > #define SCHED_CPUFREQ_DL (1U << 1) > #define SCHED_CPUFREQ_IOWAIT (1U << 2) > +#define SCHED_CPUFREQ_IDLE (1U << 3) > > #define SCHED_CPUFREQ_RT_DL (SCHED_CPUFREQ_RT | SCHED_CPUFREQ_DL) > > diff --git a/kernel/sched/cpufreq_schedutil.c b/kernel/sched/cpufreq_schedutil.c > index eaba6d6..004ae18 100644 > --- a/kernel/sched/cpufreq_schedutil.c > +++ b/kernel/sched/cpufreq_schedutil.c > @@ -304,6 +304,12 @@ static void sugov_update_shared(struct update_util_data *hook, u64 time, > > sg_cpu->util = util; > sg_cpu->max = max; > + > + /* CPU is entering IDLE, reset flags without triggering an update */ > + if (unlikely(flags & SCHED_CPUFREQ_IDLE)) { > + sg_cpu->flags = 0; > + goto done; > + } Instead of defining a new flag for idle, wouldn't another way be to just clear the flag from the RT scheduling class with an extra call to cpufreq_update_util with flags = 0 during dequeue_rt_entity? That seems to me to be also the right place to clear the flag since the flag is set in the corresponding class to begin with. thanks, -Joel
On Tue, Jul 4, 2017 at 10:34 AM, Patrick Bellasi <patrick.bellasi@arm.com> wrote: > Currently the utilization of the FAIR class is collected before locking > the policy. Although that should not be a big issue for most cases, we > also don't really know how much latency there can be between the > utilization reading and its usage. > > Let's get the FAIR utilization right before its usage to be better in > sync with the current status of a CPU. > > Signed-off-by: Patrick Bellasi <patrick.bellasi@arm.com> > Cc: Ingo Molnar <mingo@redhat.com> > Cc: Peter Zijlstra <peterz@infradead.org> > Cc: Rafael J. Wysocki <rafael.j.wysocki@intel.com> > Cc: Viresh Kumar <viresh.kumar@linaro.org> > Cc: linux-kernel@vger.kernel.org > Cc: linux-pm@vger.kernel.org > --- > kernel/sched/cpufreq_schedutil.c | 3 +-- > 1 file changed, 1 insertion(+), 2 deletions(-) > > diff --git a/kernel/sched/cpufreq_schedutil.c b/kernel/sched/cpufreq_schedutil.c > index 98704d8..df433f1 100644 > --- a/kernel/sched/cpufreq_schedutil.c > +++ b/kernel/sched/cpufreq_schedutil.c > @@ -308,10 +308,9 @@ static void sugov_update_shared(struct update_util_data *hook, u64 time, > if (unlikely(current == sg_policy->thread)) > return; > > - sugov_get_util(&util, &max); > - > raw_spin_lock(&sg_policy->update_lock); > > + sugov_get_util(&util, &max); > sg_cpu->util = util; > sg_cpu->max = max; Is your concern that there will we spinlock contention before calling sugov_get_util? If that's the case, with your patch it seems to me such contention (and hence spinning) itself could cause the utilization to be inflated - thus calling sugov_get_util after acquiring the lock will not be as accurate as before. In that case it seems to me its better to let sugov_get_util be called before acquiring the lock (as before). thanks, -Joel
On 2017-07-04 10:34, Patrick Bellasi wrote: > Currently the utilization of the FAIR class is collected before locking > the policy. Although that should not be a big issue for most cases, we > also don't really know how much latency there can be between the > utilization reading and its usage. > > Let's get the FAIR utilization right before its usage to be better in > sync with the current status of a CPU. > > Signed-off-by: Patrick Bellasi <patrick.bellasi@arm.com> > Cc: Ingo Molnar <mingo@redhat.com> > Cc: Peter Zijlstra <peterz@infradead.org> > Cc: Rafael J. Wysocki <rafael.j.wysocki@intel.com> > Cc: Viresh Kumar <viresh.kumar@linaro.org> > Cc: linux-kernel@vger.kernel.org > Cc: linux-pm@vger.kernel.org > --- > kernel/sched/cpufreq_schedutil.c | 3 +-- > 1 file changed, 1 insertion(+), 2 deletions(-) > > diff --git a/kernel/sched/cpufreq_schedutil.c > b/kernel/sched/cpufreq_schedutil.c > index 98704d8..df433f1 100644 > --- a/kernel/sched/cpufreq_schedutil.c > +++ b/kernel/sched/cpufreq_schedutil.c > @@ -308,10 +308,9 @@ static void sugov_update_shared(struct > update_util_data *hook, u64 time, > if (unlikely(current == sg_policy->thread)) > return; > > - sugov_get_util(&util, &max); > - > raw_spin_lock(&sg_policy->update_lock); > > + sugov_get_util(&util, &max); > sg_cpu->util = util; > sg_cpu->max = max; Given that the utilization update hooks are called with the per-cpu rq lock held (for all classes), I don't think PELT utilization can change throughout the lifetime of the cpufreq_update_{util,this_cpu} call? Even with Viresh's remote cpu callback series we'd still have to hold the rq lock across cpufreq_update_util.. what can change today is 'max' (arch_scale_cpu_capacity) when a cpufreq policy is shared, so the patch is still needed for that reason I think? Thanks, Vikram -- Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum, a Linux Foundation Collaborative Project
On 06/07/17 21:43, Joel Fernandes wrote: > On Tue, Jul 4, 2017 at 10:34 AM, Patrick Bellasi > <patrick.bellasi@arm.com> wrote: [...] > > @@ -304,6 +304,12 @@ static void sugov_update_shared(struct update_util_data *hook, u64 time, > > > > sg_cpu->util = util; > > sg_cpu->max = max; > > + > > + /* CPU is entering IDLE, reset flags without triggering an update */ > > + if (unlikely(flags & SCHED_CPUFREQ_IDLE)) { > > + sg_cpu->flags = 0; > > + goto done; > > + } > > Instead of defining a new flag for idle, wouldn't another way be to > just clear the flag from the RT scheduling class with an extra call to > cpufreq_update_util with flags = 0 during dequeue_rt_entity? That > seems to me to be also the right place to clear the flag since the > flag is set in the corresponding class to begin with. > Make sense to me too. Also considering that for DL (with my patches) we don't generally want to clear the flag at dequeue time, but only when the 0-lag timer fires. Best, - Juri
Hi Vikram, On Thu, Jul 6, 2017 at 11:44 PM, Vikram Mulukutla <markivx@codeaurora.org> wrote: > On 2017-07-04 10:34, Patrick Bellasi wrote: >> >> Currently the utilization of the FAIR class is collected before locking >> the policy. Although that should not be a big issue for most cases, we >> also don't really know how much latency there can be between the >> utilization reading and its usage. >> >> Let's get the FAIR utilization right before its usage to be better in >> sync with the current status of a CPU. >> >> Signed-off-by: Patrick Bellasi <patrick.bellasi@arm.com> >> Cc: Ingo Molnar <mingo@redhat.com> >> Cc: Peter Zijlstra <peterz@infradead.org> >> Cc: Rafael J. Wysocki <rafael.j.wysocki@intel.com> >> Cc: Viresh Kumar <viresh.kumar@linaro.org> >> Cc: linux-kernel@vger.kernel.org >> Cc: linux-pm@vger.kernel.org >> --- >> kernel/sched/cpufreq_schedutil.c | 3 +-- >> 1 file changed, 1 insertion(+), 2 deletions(-) >> >> diff --git a/kernel/sched/cpufreq_schedutil.c >> b/kernel/sched/cpufreq_schedutil.c >> index 98704d8..df433f1 100644 >> --- a/kernel/sched/cpufreq_schedutil.c >> +++ b/kernel/sched/cpufreq_schedutil.c >> @@ -308,10 +308,9 @@ static void sugov_update_shared(struct >> update_util_data *hook, u64 time, >> if (unlikely(current == sg_policy->thread)) >> return; >> >> - sugov_get_util(&util, &max); >> - >> raw_spin_lock(&sg_policy->update_lock); >> >> + sugov_get_util(&util, &max); >> sg_cpu->util = util; >> sg_cpu->max = max; > > > Given that the utilization update hooks are called with the per-cpu rq lock > held (for all classes), I don't think PELT utilization can change throughout > the lifetime of the cpufreq_update_{util,this_cpu} call? Even with Viresh's > remote cpu callback series we'd still have to hold the rq lock across > cpufreq_update_util.. what can change today is 'max' > (arch_scale_cpu_capacity) when a cpufreq policy is shared, so the patch is > still needed for that reason I think? > I didn't follow, Could you elaborate more why you think the patch helps with the case where max changes while the per-cpu rq lock held? thanks, -Joel
On 2017-07-07 23:14, Joel Fernandes wrote: > Hi Vikram, > > On Thu, Jul 6, 2017 at 11:44 PM, Vikram Mulukutla > <markivx@codeaurora.org> wrote: >> On 2017-07-04 10:34, Patrick Bellasi wrote: >>> >>> Currently the utilization of the FAIR class is collected before >>> locking >>> the policy. Although that should not be a big issue for most cases, >>> we >>> also don't really know how much latency there can be between the >>> utilization reading and its usage. >>> >>> Let's get the FAIR utilization right before its usage to be better in >>> sync with the current status of a CPU. >>> >>> Signed-off-by: Patrick Bellasi <patrick.bellasi@arm.com> <snip> >> Given that the utilization update hooks are called with the per-cpu rq >> lock >> held (for all classes), I don't think PELT utilization can change >> throughout >> the lifetime of the cpufreq_update_{util,this_cpu} call? Even with >> Viresh's >> remote cpu callback series we'd still have to hold the rq lock across >> cpufreq_update_util.. what can change today is 'max' >> (arch_scale_cpu_capacity) when a cpufreq policy is shared, so the >> patch is >> still needed for that reason I think? >> > > I didn't follow, Could you elaborate more why you think the patch > helps with the case where max changes while the per-cpu rq lock held? > So going by Patrick's commit text, the concern was a TOC/TOU problem, but since we agree that CFS utilization can't change within an rq-locked critical section, the only thing that can change is 'max'. So you might be the 8th cpu in line waiting for the sg-policy lock, and after you get the lock, the max may be outdated. But come to think of it max changes should be triggering schedutil updates and those shouldn't be rate-throttled, so maybe we don't need this at all? It's still somewhat future-proof in case there is some stat that we read in sugov_get_util that can be updated asynchronously. However we can put it in when we need it... > thanks, > > -Joel -- Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum, a Linux Foundation Collaborative Project
On Mon, Jul 10, 2017 at 10:49 AM, Vikram Mulukutla <markivx@codeaurora.org> wrote: [..] > >>> Given that the utilization update hooks are called with the per-cpu rq >>> lock >>> held (for all classes), I don't think PELT utilization can change >>> throughout >>> the lifetime of the cpufreq_update_{util,this_cpu} call? Even with >>> Viresh's >>> remote cpu callback series we'd still have to hold the rq lock across >>> cpufreq_update_util.. what can change today is 'max' >>> (arch_scale_cpu_capacity) when a cpufreq policy is shared, so the patch >>> is >>> still needed for that reason I think? >>> >> >> I didn't follow, Could you elaborate more why you think the patch >> helps with the case where max changes while the per-cpu rq lock held? >> > > So going by Patrick's commit text, the concern was a TOC/TOU > problem, but since we agree that CFS utilization can't change > within an rq-locked critical section, the only thing that can > change is 'max'. So you might be the 8th cpu in line waiting > for the sg-policy lock, and after you get the lock, the max may > be outdated. > > But come to think of it max changes should be triggering schedutil > updates and those shouldn't be rate-throttled, so maybe we don't > need this at all? It's still somewhat future-proof in case there > is some stat that we read in sugov_get_util that can be updated > asynchronously. However we can put it in when we need it... It looks like Juri's patch [1] to split signals already cleaned it up so we should be all set ;-) Thanks, -Joel [1] https://patchwork.kernel.org/patch/9826201/
On 07/04/2017 10:34 AM, Patrick Bellasi wrote: > In system where multiple CPUs shares the same frequency domain a small > workload on a CPU can still be subject to frequency spikes, generated by > the activation of the sugov's kthread. > > Since the sugov kthread is a special RT task, which goal is just that to > activate a frequency transition, it does not make sense for it to bias > the schedutil's frequency selection policy. > > This patch exploits the information related to the current task to silently > ignore cpufreq_update_this_cpu() calls, coming from the RT scheduler, while > the sugov kthread is running. > > Signed-off-by: Patrick Bellasi <patrick.bellasi@arm.com> > Cc: Ingo Molnar <mingo@redhat.com> > Cc: Peter Zijlstra <peterz@infradead.org> > Cc: Rafael J. Wysocki <rafael.j.wysocki@intel.com> > Cc: Viresh Kumar <viresh.kumar@linaro.org> > Cc: linux-kernel@vger.kernel.org > Cc: linux-pm@vger.kernel.org > > --- > Changes from v1: > - move check before policy spinlock (JuriL) > --- > kernel/sched/cpufreq_schedutil.c | 8 ++++++++ > 1 file changed, 8 insertions(+) > > diff --git a/kernel/sched/cpufreq_schedutil.c b/kernel/sched/cpufreq_schedutil.c > index c982dd0..eaba6d6 100644 > --- a/kernel/sched/cpufreq_schedutil.c > +++ b/kernel/sched/cpufreq_schedutil.c > @@ -218,6 +218,10 @@ static void sugov_update_single(struct update_util_data *hook, u64 time, > unsigned int next_f; > bool busy; > > + /* Skip updates generated by sugov kthreads */ > + if (unlikely(current == sg_policy->thread)) > + return; > + > sugov_set_iowait_boost(sg_cpu, time, flags); > sg_cpu->last_update = time; > > @@ -290,6 +294,10 @@ static void sugov_update_shared(struct update_util_data *hook, u64 time, > unsigned long util, max; > unsigned int next_f; > > + /* Skip updates generated by sugov kthreads */ > + if (unlikely(current == sg_policy->thread)) > + return; > + This seems super race-y. Especially when combined with rate_limit_us. Deciding to not update the frequency for a policy just because the call back happened in the context of the kthread is not right. Especially when it's combined with the remote CPU call backs patches Viresh is putting out (which I think is a well intended patch series). -Saravana -- Qualcomm Innovation Center, Inc. The Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum, a Linux Foundation Collaborative Project
On 07/07/2017 03:17 AM, Juri Lelli wrote: > On 06/07/17 21:43, Joel Fernandes wrote: >> On Tue, Jul 4, 2017 at 10:34 AM, Patrick Bellasi >> <patrick.bellasi@arm.com> wrote: > > [...] > >>> @@ -304,6 +304,12 @@ static void sugov_update_shared(struct update_util_data *hook, u64 time, >>> >>> sg_cpu->util = util; >>> sg_cpu->max = max; >>> + >>> + /* CPU is entering IDLE, reset flags without triggering an update */ >>> + if (unlikely(flags & SCHED_CPUFREQ_IDLE)) { >>> + sg_cpu->flags = 0; >>> + goto done; >>> + } >> >> Instead of defining a new flag for idle, wouldn't another way be to >> just clear the flag from the RT scheduling class with an extra call to >> cpufreq_update_util with flags = 0 during dequeue_rt_entity? That >> seems to me to be also the right place to clear the flag since the >> flag is set in the corresponding class to begin with. >> > > Make sense to me too. Also considering that for DL (with my patches) we > don't generally want to clear the flag at dequeue time, but only when > the 0-lag timer fires. > Makes sense to me too. -Saravana -- Qualcomm Innovation Center, Inc. The Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum, a Linux Foundation Collaborative Project
Each time a CPU utilisation update is issued by the scheduler a flag, which mainly defines which scheduling class is asking for the update, is used by the frequency selection policy to support the selection of the most appropriate OPP. In the current implementation, CPU flags are overridden each time the scheduler calls schedutil for an update. Such a behavior seems to be sub-optimal, especially on systems where frequency domains span across multiple CPUs. Indeed, assuming CPU1 and CPU2 share the same frequency domain, there can be the following issues: A) Small FAIR task running at MAX OPP. A RT task, which just executed on CPU1, can keep the domain at the max frequency for a prolonged period of time after its completion, even if there are no longer RT tasks running on CPUs of its domain. B) FAIR wakeup reducing the OPP of the current RT task. A FAIR task enqueued in a CPU where a RT task is running overrides the flag configured by the RT task thus potentially causing an unwanted frequency drop. C) RT wakeup not running at max OPP. An RT task waking up on a CPU which has recently updated its OPP can be forced to run at a lower frequency because of the throttling enforced by schedutil, even if there are not OPP transitions currently in progress. .:: Patches organization ======================== This series proposes a set of fixes for the aforementioned issues and it's an update addressing all the main comments collected from the previous posting [1]. Patches have been re-ordered to have the "less controversial" bits at the beginning and also to better match the order of the three main issues described above. These are the relative patches: A) Fix small FAIR task running at MAX OPP: cpufreq: schedutil: ignore the sugov kthread for frequencies selections cpufreq: schedutil: reset sg_cpus's flags at IDLE enter B) FAIR wakeup reducing the OPP of the current RT task. cpufreq: schedutil: ensure max frequency while running RT/DL tasks C) RT wakeup not running at max OPP. sched/rt: fast switch to maximum frequency when RT tasks are scheduled cpufreq: schedutil: relax rate-limiting while running RT/DL tasks cpufreq: schedutil: avoid utilisation update when not necessary .:: Experimental Results ======================== The misbehavior have been verified using a set of simple rt-app based synthetic workloads, running on a ARM's Juno R2 board where the CPUs of the big cluster (CPU1 and CPU2) have been reserved to run the workload tasks in isolation from other system tasks. A detailed description of the experiments executed, and the corresponding collected results, is available [2] online. Short highlights for these experiments are: - Patches in group A reduce energy consumption by ~50% by ensuring that a small task is always running at the minimum OPP even when the sugov's RT kthread is used to change frequencies in the same cluster. - Patches in group B increase from 4% to 98% the chances for a RT task to complete its activations while running at the max OPP. - Patches in group C do not show measurable differences mainly because of the slow OPP switching support available on the JUNO board used for testing. However, a trace inspection shows that the sequence of traced events is much more deterministic and it better matches the expected system behaviors. For example, as soon as a RT task wakeup the scheduler ask for an OPP switch to max frequency. Cheers Patrick .:: References ============== [1] https://lkml.org/lkml/2017/3/2/385 [2] https://gist.github.com/derkling/0cd7210e4fa6f2ec3558073006e5ad70 Patrick Bellasi (6): cpufreq: schedutil: ignore sugov kthreads cpufreq: schedutil: reset sg_cpus's flags at IDLE enter cpufreq: schedutil: ensure max frequency while running RT/DL tasks cpufreq: schedutil: update CFS util only if used sched/rt: fast switch to maximum frequency when RT tasks are scheduled cpufreq: schedutil: relax rate-limiting while running RT/DL tasks include/linux/sched/cpufreq.h | 1 + kernel/sched/cpufreq_schedutil.c | 61 ++++++++++++++++++++++++++++++++-------- kernel/sched/idle_task.c | 4 +++ kernel/sched/rt.c | 15 ++++++++-- 4 files changed, 67 insertions(+), 14 deletions(-) -- 2.7.4