This patch does a couple of optimizations in init_amu_fie(), like early
exits from paths where we don't need to continue any further, moving the
calls to topology_scale_freq_invariant() just when we need
them, instead of at the top of the routine, and avoiding calling it for
the third time.
Signed-off-by: Viresh Kumar <[email protected]>
---
V2:
- The enable/disable dance is actually required, just made a bunch of
other optimizations to make it look better.
arch/arm64/kernel/topology.c | 22 ++++++++++++++--------
1 file changed, 14 insertions(+), 8 deletions(-)
diff --git a/arch/arm64/kernel/topology.c b/arch/arm64/kernel/topology.c
index ebadc73449f9..1ebdb667f0d1 100644
--- a/arch/arm64/kernel/topology.c
+++ b/arch/arm64/kernel/topology.c
@@ -221,7 +221,7 @@ static DEFINE_STATIC_KEY_FALSE(amu_fie_key);
static int __init init_amu_fie(void)
{
- bool invariance_status = topology_scale_freq_invariant();
+ bool prev, now;
cpumask_var_t valid_cpus;
int ret = 0;
int cpu;
@@ -249,18 +249,24 @@ static int __init init_amu_fie(void)
if (cpumask_equal(valid_cpus, cpu_present_mask))
cpumask_copy(amu_fie_cpus, cpu_present_mask);
- if (!cpumask_empty(amu_fie_cpus)) {
- pr_info("CPUs[%*pbl]: counters will be used for FIE.",
- cpumask_pr_args(amu_fie_cpus));
- static_branch_enable(&amu_fie_key);
- }
+ if (cpumask_empty(amu_fie_cpus))
+ goto free_valid_mask;
+
+ prev = topology_scale_freq_invariant();
+ static_branch_enable(&amu_fie_key);
+ now = topology_scale_freq_invariant();
/*
* If the system is not fully invariant after AMU init, disable
* partial use of counters for frequency invariance.
*/
- if (!topology_scale_freq_invariant())
+ if (!now) {
static_branch_disable(&amu_fie_key);
+ goto free_valid_mask;
+ }
+
+ pr_info("CPUs[%*pbl]: counters will be used for FIE.",
+ cpumask_pr_args(amu_fie_cpus));
/*
* Task scheduler behavior depends on frequency invariance support,
@@ -268,7 +274,7 @@ static int __init init_amu_fie(void)
* a result of counter initialisation and use, retrigger the build of
* scheduling domains to ensure the information is propagated properly.
*/
- if (invariance_status != topology_scale_freq_invariant())
+ if (prev != now)
rebuild_sched_domains_energy();
free_valid_mask:
--
2.25.0.rc1.19.g042ed3e048af
Hey,
On Thursday 10 Dec 2020 at 21:59:23 (+0530), Viresh Kumar wrote:
> This patch does a couple of optimizations in init_amu_fie(), like early
> exits from paths where we don't need to continue any further, moving the
> calls to topology_scale_freq_invariant() just when we need
> them, instead of at the top of the routine, and avoiding calling it for
> the third time.
>
> Signed-off-by: Viresh Kumar <[email protected]>
> ---
> V2:
> - The enable/disable dance is actually required, just made a bunch of
> other optimizations to make it look better.
>
> arch/arm64/kernel/topology.c | 22 ++++++++++++++--------
> 1 file changed, 14 insertions(+), 8 deletions(-)
>
> diff --git a/arch/arm64/kernel/topology.c b/arch/arm64/kernel/topology.c
> index ebadc73449f9..1ebdb667f0d1 100644
> --- a/arch/arm64/kernel/topology.c
> +++ b/arch/arm64/kernel/topology.c
> @@ -221,7 +221,7 @@ static DEFINE_STATIC_KEY_FALSE(amu_fie_key);
>
> static int __init init_amu_fie(void)
> {
> - bool invariance_status = topology_scale_freq_invariant();
> + bool prev, now;
Nit: can you move this below valid_cpus? It makes the section nicer if
they are in decreasing order of line length.
> cpumask_var_t valid_cpus;
> int ret = 0;
> int cpu;
> @@ -249,18 +249,24 @@ static int __init init_amu_fie(void)
> if (cpumask_equal(valid_cpus, cpu_present_mask))
> cpumask_copy(amu_fie_cpus, cpu_present_mask);
>
> - if (!cpumask_empty(amu_fie_cpus)) {
> - pr_info("CPUs[%*pbl]: counters will be used for FIE.",
> - cpumask_pr_args(amu_fie_cpus));
> - static_branch_enable(&amu_fie_key);
> - }
> + if (cpumask_empty(amu_fie_cpus))
> + goto free_valid_mask;
> +
> + prev = topology_scale_freq_invariant();
> + static_branch_enable(&amu_fie_key);
I think there could be a potential problem here (it would be unlikely
but why not fix it :) ). It was in the code before your changes.
When we enable amu_fie_key here, topology_scale_freq_tick() could be
called for AMU CPUs, which will compute and set a scale factor. Later
on, if we happen to find the system not invariant, we disable counter
based invariance, but a scale factor might have been set already for a
CPU, which would and should have returned 1024 otherwise (the
initialisation value of freq_scale).
Therefore, while here, you could instead do the following:
cpufreq_inv = cpufreq_supports_freq_invariance();
if (!cpufreq_inv &&
!cpumask_subset(cpu_online_mask, amu_fie_cpus))
goto free_valid_mask;
static_branch_enable(&amu_fie_key);
pr_info(..);
if (!cpufreq_inv)
rebuild_sched_domains_energy();
What do you think?
I can submit this separately, if you don't want the hassle.
Thanks,
Ionela.
> + now = topology_scale_freq_invariant();
>
> /*
> * If the system is not fully invariant after AMU init, disable
> * partial use of counters for frequency invariance.
> */
> - if (!topology_scale_freq_invariant())
> + if (!now) {
> static_branch_disable(&amu_fie_key);
> + goto free_valid_mask;
> + }
> +
> + pr_info("CPUs[%*pbl]: counters will be used for FIE.",
> + cpumask_pr_args(amu_fie_cpus));
>
> /*
> * Task scheduler behavior depends on frequency invariance support,
> @@ -268,7 +274,7 @@ static int __init init_amu_fie(void)
> * a result of counter initialisation and use, retrigger the build of
> * scheduling domains to ensure the information is propagated properly.
> */
> - if (invariance_status != topology_scale_freq_invariant())
> + if (prev != now)
> rebuild_sched_domains_energy();
>
> free_valid_mask:
> --
> 2.25.0.rc1.19.g042ed3e048af
>
On 14-12-20, 14:00, Ionela Voinescu wrote:
> I think there could be a potential problem here (it would be unlikely
> but why not fix it :) ). It was in the code before your changes.
>
> When we enable amu_fie_key here, topology_scale_freq_tick() could be
> called for AMU CPUs, which will compute and set a scale factor. Later
> on, if we happen to find the system not invariant, we disable counter
> based invariance, but a scale factor might have been set already for a
> CPU, which would and should have returned 1024 otherwise (the
> initialisation value of freq_scale).
>
>
> Therefore, while here, you could instead do the following:
>
> cpufreq_inv = cpufreq_supports_freq_invariance();
>
> if (!cpufreq_inv &&
> !cpumask_subset(cpu_online_mask, amu_fie_cpus))
> goto free_valid_mask;
>
> static_branch_enable(&amu_fie_key);
>
> pr_info(..);
>
> if (!cpufreq_inv)
> rebuild_sched_domains_energy();
>
> What do you think?
I already had a patch for this, but for a different reason, i.e. to
avoid the enable/disable dance.
/* We aren't fully invariant yet */
if (!prev && !cpumask_equal(amu_fie_cpus, cpu_present_mask))
return;
And this is quite similar to what you have here.
--
viresh