Messages in this thread | | | From | Vincent Guittot <> | Date | Fri, 15 Nov 2019 16:31:35 +0100 | Subject | Re: [PATCH v4] sched/freq: move call to cpufreq_update_util |
| |
On Fri, 15 Nov 2019 at 16:12, Peter Zijlstra <peterz@infradead.org> wrote: > > On Fri, Nov 15, 2019 at 02:37:27PM +0100, Vincent Guittot wrote: > > On Fri, 15 Nov 2019 at 14:25, Peter Zijlstra <peterz@infradead.org> wrote: > > > > Should not all 3 have their windows aligned and thus alway return the > > > exact same value? > > > > rt and dl yes but not irq > > > > But having aligned window doesn't mean that they will all decay. > > One can have been updated just before (during a dequeue as an example) > > or at least less than 1ms before > > Now, the thing is, if that update happened in sched/rt, then it wouldn't > have called cpufreq anyway. And once we're idle longer than a period, > they'll all decay at once. > > Except indeed that IRQ stuff, which runs out of sync. > > That is, I'm just not convinced it matters much if we keep rq->cfs > on the list forever (like UP). Because we'll only stop calling when > update_blocked_averages() when everything hit 0, and up until that > point, we'll get one update per period from rq->cfs. > > For good measure we can force an update when @done, at that point we > know all 0s. > > How is something like this? > > --- > diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c > index 545bcb90b4de..a99ac2aa4a23 100644 > --- a/kernel/sched/fair.c > +++ b/kernel/sched/fair.c > @@ -3508,9 +3508,6 @@ update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq) > cfs_rq->load_last_update_time_copy = sa->last_update_time; > #endif > > - if (decayed) > - cfs_rq_util_change(cfs_rq, 0); > - > return decayed; > } > > @@ -3620,8 +3617,12 @@ static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *s > attach_entity_load_avg(cfs_rq, se, SCHED_CPUFREQ_MIGRATION); > update_tg_load_avg(cfs_rq, 0); > > - } else if (decayed && (flags & UPDATE_TG)) > - update_tg_load_avg(cfs_rq, 0); > + } else if (decayed) { > + cfs_rq_util_change(cfs_rq, 0); > + > + if (flags & UPDATE_TG) > + update_tg_load_avg(cfs_rq, 0); > + } > } > > #ifndef CONFIG_64BIT > @@ -7453,7 +7454,7 @@ static void update_blocked_averages(int cpu) > struct cfs_rq *cfs_rq, *pos; > const struct sched_class *curr_class; > struct rq_flags rf; > - bool done = true; > + bool done = true, decayed = false; > > rq_lock_irqsave(rq, &rf); > update_rq_clock(rq); > @@ -7476,10 +7477,14 @@ static void update_blocked_averages(int cpu) > * list_add_leaf_cfs_rq() for details. > */ > for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) { > + bool last = cfs_rq == &rq->cfs; > struct sched_entity *se; > > - if (update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq), cfs_rq)) > + if (update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq), cfs_rq)) { > update_tg_load_avg(cfs_rq, 0); > + if (last)
using this last make code more readable
> + decayed = true; > + } > > /* Propagate pending load changes to the parent, if any: */ > se = cfs_rq->tg->se[cpu]; > @@ -7490,7 +7495,7 @@ static void update_blocked_averages(int cpu) > * There can be a lot of idle CPU cgroups. Don't let fully > * decayed cfs_rqs linger on the list. > */ > - if (cfs_rq_is_decayed(cfs_rq)) > + if (!last && cfs_rq_is_decayed(cfs_rq)) > list_del_leaf_cfs_rq(cfs_rq);
Keeping root cfs in the list will not change anything now that cfs_rq_util_change is in update_load_avg() cfs_rq_util_change will not be called
> > /* Don't need periodic decay once load/util_avg are null */ > @@ -7498,6 +7503,9 @@ static void update_blocked_averages(int cpu) > done = false; > } > > + if (decayed || done)
I'm not sure to get why you want to call cpufreq when done is true which means that everything reaches 0 Why do you prefer to use done instead of ORing the decay of rt, dl, irq and cfs ?
> + cpufreq_update_util(rq, 0); > + > update_blocked_load_status(rq, !done); > rq_unlock_irqrestore(rq, &rf); > } > @@ -7555,6 +7563,7 @@ static inline void update_blocked_averages(int cpu) > struct cfs_rq *cfs_rq = &rq->cfs; > const struct sched_class *curr_class; > struct rq_flags rf; > + bool done, decayed; > > rq_lock_irqsave(rq, &rf); > update_rq_clock(rq); > @@ -7568,9 +7577,13 @@ static inline void update_blocked_averages(int cpu) > update_dl_rq_load_avg(rq_clock_pelt(rq), rq, curr_class == &dl_sched_class); > update_irq_load_avg(rq, 0); > > - update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq), cfs_rq); > + decayed = update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq), cfs_rq); > + done = !(cfs_rq_has_blocked(cfs_rq) || others_have_blocked(rq)); > > - update_blocked_load_status(rq, cfs_rq_has_blocked(cfs_rq) || others_have_blocked(rq)); > + if (decayed || done) > + cpufreq_update_util(rq, 0); > + > + update_blocked_load_status(rq, !done); > rq_unlock_irqrestore(rq, &rf); > } >
| |