Skip to contentSkip to conversation

cpufreq: schedutil: Don't consider freq reduction to busy CPU if need_freq_update is set

PATCH Viresh Kumar 4 participants 14 messages Feb 18, 2021 – Feb 25, 2021 1 message not in the archive

Subsystems (2)
lore
Share & tools
lore

A message that is not in the archive (sent off-list or not yet received): [email protected]

Viresh Kumar

author maintainer
#1 lore raw
Quoted 6 lines · Yue Hu

From: Yue Hu <[email protected]>

For busy CPU case, we do not need to avoid freq reduction if limits change since commit 600f5badb78c ("cpufreq: schedutil: Don't skip freq update when limits change").

Quoted 7 lines

Later, commit 23a881852f3e ("cpufreq: schedutil: Don't skip freq update if need_freq_update is set") discarded the need_freq_update check for special case of busy CPU because we won't abort a frequency update anymore if need_freq_update is set.

That is nonlogical since we will not reduce the freq for busy CPU if the computed next_f is really reduced when limits change.

Schedutil governor will probably ask for a higher frequency than allowed, but cpufreq core will clamp the request between policy min/max before updating the frequency here.

We added the check in 600f5badb78c here earlier as there were chances that we will abort the operation without reaching to cpufreq core, which won't happen now.

Signature · 1 line
viresh
#2 lore raw
Quoted 15 lines · On Thu, 18 Feb 2021 15:50:29 +0530 Viresh Kumar <[email protected]> wrote:
Quoted 6 lines · Yue Hu

From: Yue Hu <[email protected]>

For busy CPU case, we do not need to avoid freq reduction if limits change since commit 600f5badb78c ("cpufreq: schedutil: Don't skip freq update when limits change").

Quoted 7 lines

Later, commit 23a881852f3e ("cpufreq: schedutil: Don't skip freq update if need_freq_update is set") discarded the need_freq_update check for special case of busy CPU because we won't abort a frequency update anymore if need_freq_update is set.

That is nonlogical since we will not reduce the freq for busy CPU if the computed next_f is really reduced when limits change.

Quoted 8 lines

Schedutil governor will probably ask for a higher frequency than allowed, but cpufreq core will clamp the request between policy min/max before updating the frequency here.

We added the check in 600f5badb78c here earlier as there were chances that we will abort the operation without reaching to cpufreq core, which won't happen now.

There's a possibility: we will use the previous freq to update if next_f is reduced for busy CPU if need_freq_update is set in sugov_update_next_freq(). This possibility would happen now? And this update is what we want if it happens? This is related to another possible patch ready to send.

Thank you.

author maintainer
#3 lore raw
Quoted 3 lines · Yue Hu

There's a possibility: we will use the previous freq to update if next_f is reduced for busy CPU if need_freq_update is set in sugov_update_next_freq().

Right.

Quoted 2 lines

This possibility would happen now? And this
update is what we want if it happens?

This is exactly what we want here, don't reduce speed for busy CPU, but we also need to make sure we are in the policy's valid range which cpufreq core will take care of.

Quoted 1 line

This is related to another possible patch ready to send.

I am not sure what's there to send now.

Signature · 1 line
viresh
#4 lore raw
Quoted 11 lines · On Fri, 19 Feb 2021 09:39:33 +0530 Viresh Kumar <[email protected]> wrote:
Quoted 3 lines · Yue Hu

There's a possibility: we will use the previous freq to update if next_f is reduced for busy CPU if need_freq_update is set in sugov_update_next_freq().

Right.

Quoted 2 lines

This possibility would happen now? And this
update is what we want if it happens?

This is exactly what we want here, don't reduce speed for busy CPU,

I understand it should not skip this update but set the same freq as previous one again for the specail case if need_freq_update is set. Am i rt?

Quoted 6 lines

but we also need to make sure we are in the policy's valid range which cpufreq core will take care of.

Quoted 1 line

This is related to another possible patch ready to send.

I am not sure what's there to send now.

I will send later after figure out the doubt above.

Quoted 1 line
author maintainer
#5 lore raw
Quoted 10 lines · Yue Hu
Quoted 7 lines · On Fri, 19 Feb 2021 09:39:33 +0530 Viresh Kumar <[email protected]> wrote:
Quoted 3 lines · Yue Hu

There's a possibility: we will use the previous freq to update if next_f is reduced for busy CPU if need_freq_update is set in sugov_update_next_freq().

Right.

Quoted 8 lines
Quoted 4 lines
Quoted 2 lines

This possibility would happen now? And this
update is what we want if it happens?

This is exactly what we want here, don't reduce speed for busy CPU,

I understand it should not skip this update but set the same freq as previous one again for the special case if need_freq_update is set. Am i rt?

The special check, about not reducing freq if CPU had been busy recently, doesn't have anything to do with need_freq_update.

Though previously we added the need_freq_update check there to make sure we account for any recent policy min/max change and don't skip freq update anymore. That won't happen anymore and so we don't need any check here related to need_freq_update.

If you still have doubt, please explain your concern in detail with an example as I am failing to understand it.

Signature · 1 line
viresh
#6 lore raw
Quoted 17 lines · On Fri, 19 Feb 2021 13:12:49 +0530 Viresh Kumar <[email protected]> wrote:
Quoted 16 lines · Yue Hu
Quoted 12 lines · On Fri, 19 Feb 2021 09:39:33 +0530 Viresh Kumar <[email protected]> wrote:
Quoted 3 lines · Yue Hu

There's a possibility: we will use the previous freq to update if next_f is reduced for busy CPU if need_freq_update is set in sugov_update_next_freq().

Right.

Quoted 2 lines

This possibility would happen now? And this
update is what we want if it happens?

This is exactly what we want here, don't reduce speed for busy CPU,

Quoted 6 lines
Quoted 3 lines

I understand it should not skip this update but set the same freq as previous one again for the special case if need_freq_update is set. Am i rt?

The special check, about not reducing freq if CPU had been busy recently, doesn't have anything to do with need_freq_update.

However, we will skip the update if need_freq_update is not set. And do the update if need_freq_update is set.

Note that there are unnecessary fast switch check and spin lock/unlock operations in freq skip path.

If we consider unnecessary behaviors above, then we should return right away rather than continue to execute following code.

Consequently, we also need to consider need_update_freq flag since we need to update freq currently if it is set.

Quoted 9 lines

Though previously we added the need_freq_update check there to make sure we account for any recent policy min/max change and don't skip freq update anymore. That won't happen anymore and so we don't need any check here related to need_freq_update.

If you still have doubt, please explain your concern in detail with an example as I am failing to understand it.

author maintainer
#7 lore raw
Quoted 1 line · Yue Hu

However, we will skip the update if need_freq_update is not set.

Not really, we will update freq periodically nevertheless, around every 10ms or something..

Quoted 1 line

And do the update if need_freq_update is set.

Yeah, that breaks the periodic cycle to attend to some urgent request.

Quoted 2 lines

Note that there are unnecessary fast switch check and spin lock/unlock operations in freq skip path.

Maybe, I am not sure. We are all up for optimizations if there are any.

Quoted 2 lines

If we consider unnecessary behaviors above, then we should return right away rather than continue to execute following code.

As I said earlier, we may end up updating the frequency even if need_freq_update is unset.

Signature · 1 line
viresh
#8 lore raw
Quoted 8 lines · On Fri, 19 Feb 2021 15:05:51 +0530 Viresh Kumar <[email protected]> wrote:
Quoted 1 line · Yue Hu

However, we will skip the update if need_freq_update is not set.

Not really, we will update freq periodically nevertheless, around every 10ms or something..

Quoted 1 line

And do the update if need_freq_update is set.

Quoted 7 lines

Yeah, that breaks the periodic cycle to attend to some urgent request.

Quoted 2 lines

Note that there are unnecessary fast switch check and spin lock/unlock operations in freq skip path.

Maybe, I am not sure. We are all up for optimizations if there are any.

We will set next_f to next_freq(previous freq) if next_f is reduced for busy CPU. Then the next sugov_update_next_freq() will check if next_freq matches next_f if need_freq_update is not set. Obviously, we will do nothing for the case. And The related check to fast_switch_enabled and raw_spin_{lock,unlock} operations are unnecessary.

Quoted 7 lines
Quoted 3 lines

If we consider unnecessary behaviors above, then we should return right away rather than continue to execute following code.

As I said earlier, we may end up updating the frequency even if need_freq_update is unset.

author maintainer
#9 lore raw
Quoted 6 lines · Yue Hu

We will set next_f to next_freq(previous freq) if next_f is reduced for busy CPU. Then the next sugov_update_next_freq() will check if next_freq matches next_f if need_freq_update is not set. Obviously, we will do nothing for the case. And The related check to fast_switch_enabled and raw_spin_{lock,unlock} operations are unnecessary.

Right, but we will still need sugov_update_next_freq() to have the same implementation regardless and so I am not sure if we should add this change:

kernel/sched/cpufreq_schedutil.c +3 -0
index 41e498b0008a..7289e1adab73 100644
@@ -362,6 +362,9 @@ static void sugov_update_single_freq(struct update_util_data *hook, u64 time,
         * recently, as the reduction is likely to be premature then.
         */
        if (sugov_cpu_is_busy(sg_cpu) && next_f < sg_policy->next_freq) {
+               if (!sg_policy->need_freq_update)
+                       return;
+
                next_f = sg_policy->next_freq;
 
                /* Restore cached freq as next_freq has changed */
Signature · 1 line
viresh
#10 lore raw
Quoted 10 lines · On Mon, 22 Feb 2021 11:00:14 +0530 Viresh Kumar <[email protected]> wrote:
Quoted 6 lines · Yue Hu

We will set next_f to next_freq(previous freq) if next_f is reduced for busy CPU. Then the next sugov_update_next_freq() will check if next_freq matches next_f if need_freq_update is not set. Obviously, we will do nothing for the case. And The related check to fast_switch_enabled and raw_spin_{lock,unlock} operations are unnecessary.

Right, but we will still need sugov_update_next_freq() to have the same implementation regardless and so I am not sure if we should add

Yes, sugov_update_next_freq() should be keeping current logic for corner case.

kernel/sched/cpufreq_schedutil.c static void sugov_update_single_freq(struct update_util_data *hook, u64 time, · line 365

Quoted 11 lines

this change:

kernel/sched/cpufreq_schedutil.c +1 -0
index 41e498b0008a..7289e1adab73 100644
@@ -362,6 +362,9 @@ static void sugov_update_single_freq(struct update_util_data *hook, u64 time,
         * recently, as the reduction is likely to be premature then.
         */
        if (sugov_cpu_is_busy(sg_cpu) && next_f < sg_policy->next_freq) {
+               if (!sg_policy->need_freq_update)

The initial purpose about code of next_f = sg_policy->next_freq here (for special CPU busy case) should be skipping the freq update.

Since commit 600f5badb78c ("cpufreq: schedutil: Don't skip freq update when limits change"), we add the check to busy CPU for not skipping the update, we need to update the freq using computed one because limits change.

After commit 23a881852f3e ("cpufreq: schedutil: Don't skip freq update if need_freq_update is set"), we removed the need_freq_update check(no issue of commit 600f5badb78c anymore?) and introduce to always do an update in sugov_update_next_freq() if need_freq_update is set even though current freq == sg_policy->next_freq because of corner case issue. But that is conflict with original purpose of the freq skip code (next_f = sg_policy->next_freq) of busy CPU.

kernel/sched/cpufreq_schedutil.c static void sugov_update_single_freq(struct update_util_data *hook, u64 time, · line 367

Quoted 2 lines
kernel/sched/cpufreq_schedutil.c +2 -0
+                       return;
+

Yes, it's what i want ot add for unnecessary behaviors i mentioned above. Add return here should just be for skipping update(different from corner case in commit23a881852f3e).

Quoted 5 lines
kernel/sched/cpufreq_schedutil.c +0 -0
                next_f = sg_policy->next_freq;
 
                /* Restore cached freq as next_freq has changed */
#11 lore raw
Quoted 36 lines · Yue Hu
Quoted 10 lines · On Mon, 22 Feb 2021 11:00:14 +0530 Viresh Kumar <[email protected]> wrote:
Quoted 6 lines · Yue Hu

We will set next_f to next_freq(previous freq) if next_f is reduced for busy CPU. Then the next sugov_update_next_freq() will check if next_freq matches next_f if need_freq_update is not set. Obviously, we will do nothing for the case. And The related check to fast_switch_enabled and raw_spin_{lock,unlock} operations are unnecessary.

Right, but we will still need sugov_update_next_freq() to have the same implementation regardless and so I am not sure if we should add

Yes, sugov_update_next_freq() should be keeping current logic for corner case.

kernel/sched/cpufreq_schedutil.c static void sugov_update_single_freq(struct update_util_data *hook, u64 time, · line 365

Quoted 11 lines

this change:

kernel/sched/cpufreq_schedutil.c +1 -0
index 41e498b0008a..7289e1adab73 100644
@@ -362,6 +362,9 @@ static void sugov_update_single_freq(struct update_util_data *hook, u64 time,
         * recently, as the reduction is likely to be premature then.
         */
        if (sugov_cpu_is_busy(sg_cpu) && next_f < sg_policy->next_freq) {
+               if (!sg_policy->need_freq_update)

The initial purpose about code of next_f = sg_policy->next_freq here (for special CPU busy case) should be skipping the freq update.

Since commit 600f5badb78c ("cpufreq: schedutil: Don't skip freq update when limits change"), we add the check to busy CPU for not skipping the update, we need to update the freq using computed one because limits change.

Quoted 6 lines

After commit 23a881852f3e ("cpufreq: schedutil: Don't skip freq update if need_freq_update is set"), we removed the need_freq_update check(no issue of commit 600f5badb78c anymore?) and introduce to always do an update in sugov_update_next_freq() if need_freq_update is set even though current freq == sg_policy->next_freq because of corner case issue. But that is conflict with original purpose of the freq skip code (next_f = sg_policy->next_freq) of busy CPU.

That's because we realized that it was not always a good idea to skip the update even if next_f == sg_policy->next_freq.

That's why CPUFREQ_NEED_UPDATE_LIMITS has been introduced and the current flow is a result of subsequent code rearrangements.

#12 lore raw
Quoted 44 lines · On Mon, 22 Feb 2021 15:30:34 +0100 "Rafael J. Wysocki" <[email protected]>…
Quoted 36 lines · Yue Hu
Quoted 10 lines · On Mon, 22 Feb 2021 11:00:14 +0530 Viresh Kumar <[email protected]> wrote:
Quoted 6 lines · Yue Hu

We will set next_f to next_freq(previous freq) if next_f is reduced for busy CPU. Then the next sugov_update_next_freq() will check if next_freq matches next_f if need_freq_update is not set. Obviously, we will do nothing for the case. And The related check to fast_switch_enabled and raw_spin_{lock,unlock} operations are unnecessary.

Right, but we will still need sugov_update_next_freq() to have the same implementation regardless and so I am not sure if we should add

Yes, sugov_update_next_freq() should be keeping current logic for corner case.

kernel/sched/cpufreq_schedutil.c static void sugov_update_single_freq(struct update_util_data *hook, u64 time, · line 365

Quoted 11 lines

this change:

kernel/sched/cpufreq_schedutil.c +1 -0
index 41e498b0008a..7289e1adab73 100644
@@ -362,6 +362,9 @@ static void sugov_update_single_freq(struct update_util_data *hook, u64 time,
         * recently, as the reduction is likely to be premature then.
         */
        if (sugov_cpu_is_busy(sg_cpu) && next_f < sg_policy->next_freq) {
+               if (!sg_policy->need_freq_update)  

The initial purpose about code of next_f = sg_policy->next_freq here (for special CPU busy case) should be skipping the freq update.

Since commit 600f5badb78c ("cpufreq: schedutil: Don't skip freq update when limits change"), we add the check to busy CPU for not skipping the update, we need to update the freq using computed one because limits change.

Quoted 6 lines

After commit 23a881852f3e ("cpufreq: schedutil: Don't skip freq update if need_freq_update is set"), we removed the need_freq_update check(no issue of commit 600f5badb78c anymore?) and introduce to always do an update in sugov_update_next_freq() if need_freq_update is set even though current freq == sg_policy->next_freq because of corner case issue. But that is conflict with original purpose of the freq skip code (next_f = sg_policy->next_freq) of busy CPU.

Quoted 5 lines

That's because we realized that it was not always a good idea to skip the update even if next_f == sg_policy->next_freq.

That's why CPUFREQ_NEED_UPDATE_LIMITS has been introduced and the current flow is a result of subsequent code rearrangements.

ok, care about unnecessary(should be) behaviors(fast_switch_enabled and raw_spin_{lock,unlock}) if need_freq_update is unset?

If we care, i will send another patch (which is different from above change for busy CPU).

Thank you.

#13 lore raw
Quoted 51 lines · Yue Hu
Quoted 47 lines · On Mon, 22 Feb 2021 15:30:34 +0100 "Rafael J. Wysocki" <[email protected]>…
Quoted 36 lines · Yue Hu
Quoted 10 lines · On Mon, 22 Feb 2021 11:00:14 +0530 Viresh Kumar <[email protected]> wrote:
Quoted 6 lines · Yue Hu

We will set next_f to next_freq(previous freq) if next_f is reduced for busy CPU. Then the next sugov_update_next_freq() will check if next_freq matches next_f if need_freq_update is not set. Obviously, we will do nothing for the case. And The related check to fast_switch_enabled and raw_spin_{lock,unlock} operations are unnecessary.

Right, but we will still need sugov_update_next_freq() to have the same implementation regardless and so I am not sure if we should add

Yes, sugov_update_next_freq() should be keeping current logic for corner case.

kernel/sched/cpufreq_schedutil.c static void sugov_update_single_freq(struct update_util_data *hook, u64 time, · line 365

Quoted 11 lines

this change:

kernel/sched/cpufreq_schedutil.c +1 -0
index 41e498b0008a..7289e1adab73 100644
@@ -362,6 +362,9 @@ static void sugov_update_single_freq(struct update_util_data *hook, u64 time,
         * recently, as the reduction is likely to be premature then.
         */
        if (sugov_cpu_is_busy(sg_cpu) && next_f < sg_policy->next_freq) {
+               if (!sg_policy->need_freq_update)

The initial purpose about code of next_f = sg_policy->next_freq here (for special CPU busy case) should be skipping the freq update.

Since commit 600f5badb78c ("cpufreq: schedutil: Don't skip freq update when limits change"), we add the check to busy CPU for not skipping the update, we need to update the freq using computed one because limits change.

Quoted 6 lines

After commit 23a881852f3e ("cpufreq: schedutil: Don't skip freq update if need_freq_update is set"), we removed the need_freq_update check(no issue of commit 600f5badb78c anymore?) and introduce to always do an update in sugov_update_next_freq() if need_freq_update is set even though current freq == sg_policy->next_freq because of corner case issue. But that is conflict with original purpose of the freq skip code (next_f = sg_policy->next_freq) of busy CPU.

That's because we realized that it was not always a good idea to skip the update even if next_f == sg_policy->next_freq.

Quoted 7 lines
Quoted 2 lines

That's why CPUFREQ_NEED_UPDATE_LIMITS has been introduced and the current flow is a result of subsequent code rearrangements.

ok, care about unnecessary(should be) behaviors(fast_switch_enabled and raw_spin_{lock,unlock}) if need_freq_update is unset?

If we care, i will send another patch (which is different from above change for busy CPU).

Please send a patch and we'll see (this is how things go).

#14 lore raw
Quoted 55 lines · On Wed, 24 Feb 2021 13:46:11 +0100 "Rafael J. Wysocki" <[email protected]>…
Quoted 54 lines · Yue Hu
Quoted 49 lines · On Mon, 22 Feb 2021 15:30:34 +0100 "Rafael J. Wysocki" <[email protected]>…
Quoted 36 lines · Yue Hu
Quoted 10 lines · On Mon, 22 Feb 2021 11:00:14 +0530 Viresh Kumar <[email protected]> wrote:
Quoted 6 lines · Yue Hu

We will set next_f to next_freq(previous freq) if next_f is reduced for busy CPU. Then the next sugov_update_next_freq() will check if next_freq matches next_f if need_freq_update is not set. Obviously, we will do nothing for the case. And The related check to fast_switch_enabled and raw_spin_{lock,unlock} operations are unnecessary.

Right, but we will still need sugov_update_next_freq() to have the same implementation regardless and so I am not sure if we should add

Yes, sugov_update_next_freq() should be keeping current logic for corner case.

kernel/sched/cpufreq_schedutil.c static void sugov_update_single_freq(struct update_util_data *hook, u64 time, · line 365

Quoted 11 lines

this change:

kernel/sched/cpufreq_schedutil.c +1 -0
index 41e498b0008a..7289e1adab73 100644
@@ -362,6 +362,9 @@ static void sugov_update_single_freq(struct update_util_data *hook, u64 time,
         * recently, as the reduction is likely to be premature then.
         */
        if (sugov_cpu_is_busy(sg_cpu) && next_f < sg_policy->next_freq) {
+               if (!sg_policy->need_freq_update)  

The initial purpose about code of next_f = sg_policy->next_freq here (for special CPU busy case) should be skipping the freq update.

Since commit 600f5badb78c ("cpufreq: schedutil: Don't skip freq update when limits change"), we add the check to busy CPU for not skipping the update, we need to update the freq using computed one because limits change.

Quoted 6 lines

After commit 23a881852f3e ("cpufreq: schedutil: Don't skip freq update if need_freq_update is set"), we removed the need_freq_update check(no issue of commit 600f5badb78c anymore?) and introduce to always do an update in sugov_update_next_freq() if need_freq_update is set even though current freq == sg_policy->next_freq because of corner case issue. But that is conflict with original purpose of the freq skip code (next_f = sg_policy->next_freq) of busy CPU.

That's because we realized that it was not always a good idea to skip the update even if next_f == sg_policy->next_freq.

That's why CPUFREQ_NEED_UPDATE_LIMITS has been introduced and the current flow is a result of subsequent code rearrangements.

Quoted 6 lines
Quoted 4 lines

ok, care about unnecessary(should be) behaviors(fast_switch_enabled and raw_spin_{lock,unlock}) if need_freq_update is unset?

If we care, i will send another patch (which is different from above change for busy CPU).

Please send a patch and we'll see (this is how things go).

Already sent it("Call sugov_update_next_freq() before check to fast_switch_enabled"). Please review.

Thank you.