Skip to contentSkip to conversation

Revert "xen/cpufreq: fix usages of align_timer() in the on-demand governor"

PATCH In review Jason Andryuk 4 participants 6 messages Jun 15 – Jun 16

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 Revert "xen/cpufreq: fix usages of align_timer() in the on-demand governor"

The original commit showed a ~6% regression in a benchmark. The call to align_timer(firsttick, period) rounds firsttick up to the next mutiple of the period, if firsttick % period != 0:

align_timer(0, period)          -> 0
align_timer(1, period)          -> period
align_timer(period, period)     -> period

align_timer(period + 1, period) -> 2 * period

So adding the period (sampling_rate) before calling align_timer() will in most cases incease the expiration to 2 * period (sampling_rate) (the exception being firsttick % period == 0). This longer timer slows the reaction time of the algorithm.

This reverts commit a0ed5bcfbeee81c91c574ad484faa057054eaf09.

---
This is backported in stable trees and should be reverted there as well (found in 4.20.3).

A Fixes seems superfluous and not normally used with a revert, but if needed:

Fixes a0ed5bcfbeee ("xen/cpufreq: fix usages of align_timer() in the on-demand governor")
---
 xen/drivers/cpufreq/cpufreq_ondemand.c | 5 ++---
 1 file changed, 2 insertions(+), 3 deletions(-)
xen/drivers/cpufreq/cpufreq_ondemand.c +2 -3
index 0d94c0e464..537695eaab 100644
@@ -185,8 +185,7 @@ static void cf_check do_dbs_timer(void *dbs)
     dbs_check_cpu(dbs_info);
 
     set_timer(&per_cpu(dbs_timer, dbs_info->cpu),
-              align_timer(NOW() + dbs_tuners_ins.sampling_rate,
-                          dbs_tuners_ins.sampling_rate));
+            align_timer(NOW() , dbs_tuners_ins.sampling_rate));
 }
 
 static void dbs_timer_init(struct cpu_dbs_info_s *dbs_info)
@@ -401,6 +400,6 @@ void cpufreq_dbs_timer_resume(void)
             (void)cmpxchg(stoppable, -1, 1);
         }
         else
-            set_timer(t, align_timer(t->expires, dbs_tuners_ins.sampling_rate));
+            set_timer(t, align_timer(now, dbs_tuners_ins.sampling_rate));
     }
 }
Signature · 1 line
2.54.0
#1 lore raw
Quoted 16 lines · Jason Andryuk

The original commit showed a ~6% regression in a benchmark. The call to align_timer(firsttick, period) rounds firsttick up to the next mutiple of the period, if firsttick % period != 0:

align_timer(0, period)          -> 0
align_timer(1, period)          -> period
align_timer(period, period)     -> period

align_timer(period + 1, period) -> 2 * period

So adding the period (sampling_rate) before calling align_timer() will in most cases incease the expiration to 2 * period (sampling_rate) (the exception being firsttick % period == 0). This longer timer slows the reaction time of the algorithm.

This reverts commit a0ed5bcfbeee81c91c574ad484faa057054eaf09.

Quoted 8 lines

---
This is backported in stable trees and should be reverted there as well (found in 4.20.3).

A Fixes seems superfluous and not normally used with a revert, but if needed:

Fixes a0ed5bcfbeee ("xen/cpufreq: fix usages of align_timer() in the on-demand governor")

If "git grep" is intended to (also) find such straight reverts by merely matching "Fixes: ", I think the tag should still be put there.

Talking of Fixes: tags - the original change had two of them, so there must have been a problem? Or, Roger, was this purely an observation from looking the the code?

xen/drivers/cpufreq/cpufreq_ondemand.c static void cf_check do_dbs_timer(void *dbs) · line 189

Quoted 10 lines
xen/drivers/cpufreq/cpufreq_ondemand.c +1 -2
@@ -185,8 +185,7 @@ static void cf_check do_dbs_timer(void *dbs)
     dbs_check_cpu(dbs_info);
 
     set_timer(&per_cpu(dbs_timer, dbs_info->cpu),
-              align_timer(NOW() + dbs_tuners_ins.sampling_rate,
-                          dbs_tuners_ins.sampling_rate));
+            align_timer(NOW() , dbs_tuners_ins.sampling_rate));
 }

As much as I understand you wanting to have things simple by doing a straight revert, imo the formatting flaws better wouldn't be introduced again. If I ended up committing this, I'd very likely take the liberty of doing so. Yet before ack-ing the question above needs answering.

Jan

#2 lore raw
Quoted 10 lines · Jan Beulich
Quoted 9 lines · Jason Andryuk

The original commit showed a ~6% regression in a benchmark. The call to align_timer(firsttick, period) rounds firsttick up to the next mutiple of the period, if firsttick % period != 0:

align_timer(0, period)          -> 0
align_timer(1, period)          -> period
align_timer(period, period)     -> period

align_timer(period + 1, period) -> 2 * period

Quoted 8 lines
Quoted 8 lines

So adding the period (sampling_rate) before calling align_timer() will in most cases incease the expiration to 2 * period (sampling_rate) (the exception being firsttick % period == 0). This longer timer slows the reaction time of the algorithm.

This reverts commit a0ed5bcfbeee81c91c574ad484faa057054eaf09.

Adding Oleksii for a release-ack.

Quoted 8 lines
Quoted 7 lines

---
This is backported in stable trees and should be reverted there as well (found in 4.20.3).

A Fixes seems superfluous and not normally used with a revert, but if needed:

Fixes a0ed5bcfbeee ("xen/cpufreq: fix usages of align_timer() in the on-demand governor")
Quoted 6 lines

If "git grep" is intended to (also) find such straight reverts by merely matching "Fixes: ", I think the tag should still be put there.

Talking of Fixes: tags - the original change had two of them, so there must have been a problem? Or, Roger, was this purely an observation from looking the the code?

It was purely by observation that the current timer setup will very likely fire before the set period, making the sampling intervals possibly shorter. However shorter periods are fine, we simply over-sample. Longer periods as shown by Jason have an adverse impact in the response time of the governor.

Quoted 5 lines
Quoted 5 lines
xen/drivers/cpufreq/cpufreq_ondemand.c +0 -0
@@ -185,8 +185,7 @@ static void cf_check do_dbs_timer(void *dbs)
     dbs_check_cpu(dbs_info);
 
Quoted 10 lines

xen/drivers/cpufreq/cpufreq_ondemand.c static void cf_check do_dbs_timer(void *dbs) · line 189

Quoted 5 lines
xen/drivers/cpufreq/cpufreq_ondemand.c +1 -2
     set_timer(&per_cpu(dbs_timer, dbs_info->cpu),
-              align_timer(NOW() + dbs_tuners_ins.sampling_rate,
-                          dbs_tuners_ins.sampling_rate));
+            align_timer(NOW() , dbs_tuners_ins.sampling_rate));
 }

As much as I understand you wanting to have things simple by doing a straight revert, imo the formatting flaws better wouldn't be introduced again. If I ended up committing this, I'd very likely take the liberty of doing so. Yet before ack-ing the question above needs answering.

I wouldn't mind adjusting this at commit.

Thanks, Roger.

#3 lore raw
Quoted 16 lines · Roger Pau Monné
Quoted 15 lines · Jan Beulich
Quoted 14 lines · Jason Andryuk

The original commit showed a ~6% regression in a benchmark. The call to align_timer(firsttick, period) rounds firsttick up to the next mutiple of the period, if firsttick % period != 0:

align_timer(0, period)          -> 0
align_timer(1, period)          -> period
align_timer(period, period)     -> period

align_timer(period + 1, period) -> 2 * period

So adding the period (sampling_rate) before calling align_timer() will in most cases incease the expiration to 2 * period (sampling_rate) (the exception being firsttick % period == 0). This longer timer slows the reaction time of the algorithm.

Quoted 8 lines
Quoted 3 lines
Quoted 3 lines

This reverts commit a0ed5bcfbeee81c91c574ad484faa057054eaf09.

Adding Oleksii for a release-ack.

Thanks.

~ Oleksii

#4 lore raw
Quoted 21 lines · Oleksii Kurochko
Quoted 18 lines · Roger Pau Monné
Quoted 17 lines · Jan Beulich
Quoted 16 lines · Jason Andryuk

The original commit showed a ~6% regression in a benchmark.  The call to align_timer(firsttick, period) rounds firsttick up to the next mutiple of the period, if firsttick % period != 0:

align_timer(0, period)          -> 0
align_timer(1, period)          -> period
align_timer(period, period)     -> period
align_timer(period + 1, period) -> 2 * period

So adding the period (sampling_rate) before calling align_timer() will in most cases incease the expiration to 2 * period (sampling_rate) (the exception being firsttick % period == 0).  This longer timer slows the reaction time of the algorithm.

This reverts commit a0ed5bcfbeee81c91c574ad484faa057054eaf09.

Quoted 8 lines
Quoted 6 lines
Quoted 1 line
Quoted 1 line

Adding Oleksii for a release-ack.

Thanks everyone.

Sorry for forgetting to CC you, Oleskii, and mark as for-4.22. add_maintainers.pl does the right thing... except during the code freeze.

Jan, thanks for fixing up the indent.

Regards,
Jason

#5 lore raw
In reply to Jason Andryuk
Quoted 9 lines · Jason Andryuk

The original commit showed a ~6% regression in a benchmark. The call to align_timer(firsttick, period) rounds firsttick up to the next mutiple of the period, if firsttick % period != 0:

align_timer(0, period)          -> 0
align_timer(1, period)          -> period
align_timer(period, period)     -> period

align_timer(period + 1, period) -> 2 * period

Quoted 8 lines

So adding the period (sampling_rate) before calling align_timer() will in most cases incease the expiration to 2 * period (sampling_rate) (the exception being firsttick % period == 0). This longer timer slows the reaction time of the algorithm.

This reverts commit a0ed5bcfbeee81c91c574ad484faa057054eaf09.

Oh, also Cc: Oleksii for a release-ack.

Jan