Skip to contentSkip to conversation

sched/topology: Don't set SD_BALANCE_WAKE on cpuset domain relax

PATCH Applied Valentin Schneider 3 participants 5 messages Oct 14, 2019 – Oct 18, 2019 Scheduler

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 sched/topology: Don't set SD_BALANCE_WAKE on cpuset domain relax

As pointed out in commit

182a85f8a119 ("sched: Disable wakeup balancing")

SD_BALANCE_WAKE is a tad too aggressive, and is usually left unset.

However, it turns out cpuset domain relaxation will unconditionally set it on domains below the relaxation level. This made sense back when SD_BALANCE_WAKE was set unconditionally, but it no longer is the case.

We can improve things slightly by noticing that set_domain_attribute() is always called after sd_init(), so rather than setting flags we can rely on whatever sd_init() is doing and only clear certain flags when above the relaxation level.

While at it, slightly clean up the function and flip the relax level check to be more human readable.

---
I was tempted to put a

Fixes 182a85f8a119 ("sched: Disable wakeup balancing")

but the SD setup code back then was a mess of SD_INIT() macros which I'm not familiar with. It *looks* like the sequence was roughly the same as it is now (i.e. set up domain flags, *then* call set_domain_attribute()) but I'm not completely sure.

---
 kernel/sched/topology.c | 9 +++------
 1 file changed, 3 insertions(+), 6 deletions(-)
kernel/sched/topology.c +3 -7
index b5667a273bf6..3623ffe85d18 100644
@@ -1201,17 +1201,13 @@ static void set_domain_attribute(struct sched_domain *sd,
 	if (!attr || attr->relax_domain_level < 0) {
 		if (default_relax_domain_level < 0)
 			return;
-		else
-			request = default_relax_domain_level;
+		request = default_relax_domain_level;
 	} else
 		request = attr->relax_domain_level;
-	if (request < sd->level) {
+
+	if (sd->level > request) {
 		/* Turn off idle balance on this domain: */
 		sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
-	} else {
-		/* Turn on idle balance on this domain: */
-		sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
 	}
 }
 
--

2.22.0

maintainer
#1 lore raw
Quoted 22 lines · Valentin Schneider

As pointed out in commit

182a85f8a119 ("sched: Disable wakeup balancing")

SD_BALANCE_WAKE is a tad too aggressive, and is usually left unset.

However, it turns out cpuset domain relaxation will unconditionally set it on domains below the relaxation level. This made sense back when SD_BALANCE_WAKE was set unconditionally, but it no longer is the case.

We can improve things slightly by noticing that set_domain_attribute() is always called after sd_init(), so rather than setting flags we can rely on whatever sd_init() is doing and only clear certain flags when above the relaxation level.

While at it, slightly clean up the function and flip the relax level check to be more human readable.

---
I was tempted to put a

Quoted 6 lines
Fixes 182a85f8a119 ("sched: Disable wakeup balancing")

but the SD setup code back then was a mess of SD_INIT() macros which I'm not familiar with. It *looks* like the sequence was roughly the same as it is now (i.e. set up domain flags, *then* call set_domain_attribute()) but I'm not completely sure.

This 'relax' thing is on my list of regrets. It is a terrible thing that should never have existed.

Are you actually using it or did you just stumble upon it while looking around there?

#2 lore raw
Quoted 6 lines · Peter Zijlstra

This 'relax' thing is on my list of regrets. It is a terrible thing that should never have existed.

Are you actually using it or did you just stumble upon it while looking around there?

Just staring around, I don't use cpuset stuff unless I really need to...

maintainer
#3 lore raw
Quoted 9 lines · Valentin Schneider
Quoted 6 lines · Peter Zijlstra

This 'relax' thing is on my list of regrets. It is a terrible thing that should never have existed.

Are you actually using it or did you just stumble upon it while looking around there?

Just staring around, I don't use cpuset stuff unless I really need to...

Fair enough. Applied it.

#4 lore raw
In reply to Valentin Schneider
The following commit has been merged into the sched/core branch of tip:

The following commit has been merged into the sched/core branch of tip:

Commit-ID:     9ae7ab20b4835dbea0e5fc6a5c70171dc354a72e
Gitweb:        https://git.kernel.org/tip/9ae7ab20b4835dbea0e5fc6a5c70171dc354a72e
Author:        Valentin Schneider <[email protected]>
AuthorDate:    Mon, 14 Oct 2019 17:44:08 +01:00
Committer:     Peter Zijlstra <[email protected]>
CommitterDate: Thu, 17 Oct 2019 21:31:54 +02:00

sched/topology: Don't set SD_BALANCE_WAKE on cpuset domain relax

As pointed out in commit

182a85f8a119 ("sched: Disable wakeup balancing")

SD_BALANCE_WAKE is a tad too aggressive, and is usually left unset.

However, it turns out cpuset domain relaxation will unconditionally set it on domains below the relaxation level. This made sense back when SD_BALANCE_WAKE was set unconditionally, but it no longer is the case.

We can improve things slightly by noticing that set_domain_attribute() is always called after sd_init(), so rather than setting flags we can rely on whatever sd_init() is doing and only clear certain flags when above the relaxation level.

While at it, slightly clean up the function and flip the relax level check to be more human readable.

---
 kernel/sched/topology.c |  9 +++------
 1 file changed, 3 insertions(+), 6 deletions(-)
kernel/sched/topology.c +3 -6
index b5667a2..3623ffe 100644
@@ -1201,16 +1201,13 @@ static void set_domain_attribute(struct sched_domain *sd,
 	if (!attr || attr->relax_domain_level < 0) {
 		if (default_relax_domain_level < 0)
 			return;
-		else
-			request = default_relax_domain_level;
+		request = default_relax_domain_level;
 	} else
 		request = attr->relax_domain_level;
-	if (request < sd->level) {
+
+	if (sd->level > request) {
 		/* Turn off idle balance on this domain: */
 		sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
-	} else {
-		/* Turn on idle balance on this domain: */
-		sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
 	}
 }