| Message ID | 20201016163634.857573-1-wvw@google.com (mailing list archive) |
|---|---|
| State | Superseded, archived |
| Headers | show |
| Series | sched: cpufreq_schedutil: maintain raw cache when next_f is not changed | expand |
On Fri, Oct 16, 2020 at 6:36 PM Wei Wang <wvw@google.com> wrote: > > Currently, raw cache will be reset when next_f is changed after > get_next_freq for correctness. However, it may introduce more > cycles. This patch changes it to maintain the cached value instead of > dropping it. IMV you need to be more specific about why this helps. > This is adapted from https://android-review.googlesource.com/1352810/ > > Signed-off-by: Wei Wang <wvw@google.com> > --- > kernel/sched/cpufreq_schedutil.c | 7 +++++-- > 1 file changed, 5 insertions(+), 2 deletions(-) > > diff --git a/kernel/sched/cpufreq_schedutil.c b/kernel/sched/cpufreq_schedutil.c > index 5ae7b4e6e8d6..ae3ae7fcd027 100644 > --- a/kernel/sched/cpufreq_schedutil.c > +++ b/kernel/sched/cpufreq_schedutil.c > @@ -31,6 +31,7 @@ struct sugov_policy { > s64 freq_update_delay_ns; > unsigned int next_freq; > unsigned int cached_raw_freq; > + unsigned int prev_cached_raw_freq; > > /* The next fields are only needed if fast switch cannot be used: */ > struct irq_work irq_work; > @@ -165,6 +166,7 @@ static unsigned int get_next_freq(struct sugov_policy *sg_policy, > return sg_policy->next_freq; > > sg_policy->need_freq_update = false; > + sg_policy->prev_cached_raw_freq = sg_policy->cached_raw_freq; > sg_policy->cached_raw_freq = freq; > return cpufreq_driver_resolve_freq(policy, freq); > } > @@ -464,8 +466,8 @@ static void sugov_update_single(struct update_util_data *hook, u64 time, > if (busy && next_f < sg_policy->next_freq) { > next_f = sg_policy->next_freq; > > - /* Reset cached freq as next_freq has changed */ > - sg_policy->cached_raw_freq = 0; > + /* Restore cached freq as next_freq has changed */ > + sg_policy->cached_raw_freq = sg_policy->prev_cached_raw_freq; > } > > /* > @@ -828,6 +830,7 @@ static int sugov_start(struct cpufreq_policy *policy) > sg_policy->limits_changed = false; > sg_policy->need_freq_update = false; > sg_policy->cached_raw_freq = 0; > + sg_policy->prev_cached_raw_freq = 0; > > for_each_cpu(cpu, policy->cpus) { > struct sugov_cpu *sg_cpu = &per_cpu(sugov_cpu, cpu); > -- > 2.29.0.rc1.297.gfa9743e501-goog >
On Fri, Oct 16, 2020 at 10:01 AM Rafael J. Wysocki <rafael@kernel.org> wrote: > > On Fri, Oct 16, 2020 at 6:36 PM Wei Wang <wvw@google.com> wrote: > > > > Currently, raw cache will be reset when next_f is changed after > > get_next_freq for correctness. However, it may introduce more > > cycles. This patch changes it to maintain the cached value instead of > > dropping it. > > IMV you need to be more specific about why this helps. > I think the idea of cached_raw_freq is to reduce the chance of calling cpufreq drivers (in some arch those may be costly) but sometimes the cache will be wiped for correctness. The purpose of this patch is to still keep the cached value instead of wiping them. thx wvw > > > This is adapted from https://android-review.googlesource.com/1352810/ > > > > Signed-off-by: Wei Wang <wvw@google.com> > > --- > > kernel/sched/cpufreq_schedutil.c | 7 +++++-- > > 1 file changed, 5 insertions(+), 2 deletions(-) > > > > diff --git a/kernel/sched/cpufreq_schedutil.c b/kernel/sched/cpufreq_schedutil.c > > index 5ae7b4e6e8d6..ae3ae7fcd027 100644 > > --- a/kernel/sched/cpufreq_schedutil.c > > +++ b/kernel/sched/cpufreq_schedutil.c > > @@ -31,6 +31,7 @@ struct sugov_policy { > > s64 freq_update_delay_ns; > > unsigned int next_freq; > > unsigned int cached_raw_freq; > > + unsigned int prev_cached_raw_freq; > > > > /* The next fields are only needed if fast switch cannot be used: */ > > struct irq_work irq_work; > > @@ -165,6 +166,7 @@ static unsigned int get_next_freq(struct sugov_policy *sg_policy, > > return sg_policy->next_freq; > > > > sg_policy->need_freq_update = false; > > + sg_policy->prev_cached_raw_freq = sg_policy->cached_raw_freq; > > sg_policy->cached_raw_freq = freq; > > return cpufreq_driver_resolve_freq(policy, freq); > > } > > @@ -464,8 +466,8 @@ static void sugov_update_single(struct update_util_data *hook, u64 time, > > if (busy && next_f < sg_policy->next_freq) { > > next_f = sg_policy->next_freq; > > > > - /* Reset cached freq as next_freq has changed */ > > - sg_policy->cached_raw_freq = 0; > > + /* Restore cached freq as next_freq has changed */ > > + sg_policy->cached_raw_freq = sg_policy->prev_cached_raw_freq; > > } > > > > /* > > @@ -828,6 +830,7 @@ static int sugov_start(struct cpufreq_policy *policy) > > sg_policy->limits_changed = false; > > sg_policy->need_freq_update = false; > > sg_policy->cached_raw_freq = 0; > > + sg_policy->prev_cached_raw_freq = 0; > > > > for_each_cpu(cpu, policy->cpus) { > > struct sugov_cpu *sg_cpu = &per_cpu(sugov_cpu, cpu); > > -- > > 2.29.0.rc1.297.gfa9743e501-goog > >
On Fri, Oct 16, 2020 at 7:18 PM Wei Wang <wvw@google.com> wrote: > > On Fri, Oct 16, 2020 at 10:01 AM Rafael J. Wysocki <rafael@kernel.org> wrote: > > > > On Fri, Oct 16, 2020 at 6:36 PM Wei Wang <wvw@google.com> wrote: > > > > > > Currently, raw cache will be reset when next_f is changed after > > > get_next_freq for correctness. However, it may introduce more > > > cycles. This patch changes it to maintain the cached value instead of > > > dropping it. > > > > IMV you need to be more specific about why this helps. > > > > I think the idea of cached_raw_freq is to reduce the chance of calling > cpufreq drivers (in some arch those may be costly) but sometimes the > cache will be wiped for correctness. The purpose of this patch is to > still keep the cached value instead of wiping them. Well, I see what the problem is and how the patch is attempting to address it (which is not the best way to do that because of the extra struct member that doesn't need to be added if I'm not mistaken), but IMO the changelog is way too vague from the problem statement perspective.
On Fri, Oct 16, 2020 at 10:36 AM Rafael J. Wysocki <rafael@kernel.org> wrote: > > On Fri, Oct 16, 2020 at 7:18 PM Wei Wang <wvw@google.com> wrote: > > > > On Fri, Oct 16, 2020 at 10:01 AM Rafael J. Wysocki <rafael@kernel.org> wrote: > > > > > > On Fri, Oct 16, 2020 at 6:36 PM Wei Wang <wvw@google.com> wrote: > > > > > > > > Currently, raw cache will be reset when next_f is changed after > > > > get_next_freq for correctness. However, it may introduce more > > > > cycles. This patch changes it to maintain the cached value instead of > > > > dropping it. > > > > > > IMV you need to be more specific about why this helps. > > > > > > > I think the idea of cached_raw_freq is to reduce the chance of calling > > cpufreq drivers (in some arch those may be costly) but sometimes the > > cache will be wiped for correctness. The purpose of this patch is to > > still keep the cached value instead of wiping them. > > Well, I see what the problem is and how the patch is attempting to > address it (which is not the best way to do that because of the extra > struct member that doesn't need to be added if I'm not mistaken), but > IMO the changelog is way too vague from the problem statement > perspective. Just want to bring this up in the mainline kernel. I think we can change the patch to use a variable insides sugov_update_single. This is adapted from Android common kernel where it has some off tree functions making a single variable not possible but also making the issue more obvious.
diff --git a/kernel/sched/cpufreq_schedutil.c b/kernel/sched/cpufreq_schedutil.c index 5ae7b4e6e8d6..ae3ae7fcd027 100644 --- a/kernel/sched/cpufreq_schedutil.c +++ b/kernel/sched/cpufreq_schedutil.c @@ -31,6 +31,7 @@ struct sugov_policy { s64 freq_update_delay_ns; unsigned int next_freq; unsigned int cached_raw_freq; + unsigned int prev_cached_raw_freq; /* The next fields are only needed if fast switch cannot be used: */ struct irq_work irq_work; @@ -165,6 +166,7 @@ static unsigned int get_next_freq(struct sugov_policy *sg_policy, return sg_policy->next_freq; sg_policy->need_freq_update = false; + sg_policy->prev_cached_raw_freq = sg_policy->cached_raw_freq; sg_policy->cached_raw_freq = freq; return cpufreq_driver_resolve_freq(policy, freq); } @@ -464,8 +466,8 @@ static void sugov_update_single(struct update_util_data *hook, u64 time, if (busy && next_f < sg_policy->next_freq) { next_f = sg_policy->next_freq; - /* Reset cached freq as next_freq has changed */ - sg_policy->cached_raw_freq = 0; + /* Restore cached freq as next_freq has changed */ + sg_policy->cached_raw_freq = sg_policy->prev_cached_raw_freq; } /* @@ -828,6 +830,7 @@ static int sugov_start(struct cpufreq_policy *policy) sg_policy->limits_changed = false; sg_policy->need_freq_update = false; sg_policy->cached_raw_freq = 0; + sg_policy->prev_cached_raw_freq = 0; for_each_cpu(cpu, policy->cpus) { struct sugov_cpu *sg_cpu = &per_cpu(sugov_cpu, cpu);
Currently, raw cache will be reset when next_f is changed after get_next_freq for correctness. However, it may introduce more cycles. This patch changes it to maintain the cached value instead of dropping it. This is adapted from https://android-review.googlesource.com/1352810/ Signed-off-by: Wei Wang <wvw@google.com> --- kernel/sched/cpufreq_schedutil.c | 7 +++++-- 1 file changed, 5 insertions(+), 2 deletions(-)