Skip to contentSkip to conversation

x86/irq: Do not check available vectors if current CPU has no irq to migrate

PATCH In review Chen Yu 3 participants 4 messages Apr 9, 2017 – Apr 14, 2017 x86

lore
Share & tools
lore

Chen Yu

author
#0 lore raw

PATCH 1/1 x86/irq: Do not check available vectors if current CPU has no irq to migrate

This is an optimization to bypass the free vector checking if the current CPU has no irq to migrate. This can especially speed up the CPU offline process when there are many CPUs in the system.

Cc Thomas Gleixner <[email protected]>
Cc Ingo Molnar <[email protected]>
Cc "H. Peter Anvin" <[email protected]>
Cc Prarit Bhargava <[email protected]>
Cc "Rafael J. Wysocki" <[email protected]>
Cc Len Brown <[email protected]>
---
 arch/x86/kernel/irq.c | 3 +++
 1 file changed, 3 insertions(+)
arch/x86/kernel/irq.c +3 -0
index 4d8183b..b9bdc1e 100644
@@ -394,6 +394,9 @@ int check_irq_vectors_for_cpu_disable(void)
 		    !cpumask_subset(&affinity_new, &online_new))
 			this_count++;
 	}
+	/* No need to check any further. */
+	if (!this_count)
+		return 0;
 
 	count = 0;
 	for_each_online_cpu(cpu) {
Signature · 1 line
2.7.4
#1 lore raw
Quoted 3 lines · Chen Yu

This is an optimization to bypass the free vector checking if the current CPU has no irq to migrate. This can especially speed up the CPU offline process when there are many CPUs in the system.

Chen, OOC how much time does this save?

P.

Quoted 29 lines
Cc Thomas Gleixner <[email protected]>
Cc Ingo Molnar <[email protected]>
Cc "H. Peter Anvin" <[email protected]>
Cc Prarit Bhargava <[email protected]>
Cc "Rafael J. Wysocki" <[email protected]>
Cc Len Brown <[email protected]>
---
 arch/x86/kernel/irq.c | 3 +++
 1 file changed, 3 insertions(+)
arch/x86/kernel/irq.c +3 -0
index 4d8183b..b9bdc1e 100644
@@ -394,6 +394,9 @@ int check_irq_vectors_for_cpu_disable(void)
 		    !cpumask_subset(&affinity_new, &online_new))
 			this_count++;
 	}
+	/* No need to check any further. */
+	if (!this_count)
+		return 0;
 
 	count = 0;
 	for_each_online_cpu(cpu) {
author
#2 lore raw

Hi Prarit,

Quoted 9 lines · Prarit Bhargava
Quoted 3 lines · Chen Yu

This is an optimization to bypass the free vector checking if the current CPU has no irq to migrate. This can especially speed up the CPU offline process when there are many CPUs in the system.

Chen, OOC how much time does this save?

Previously I was doing some debugging on irq related issue when check_irq_vectors_for_cpu_disable() printed a warning on that platform, thus I incidentally look at this code and did not measure it : )

I've just done a comparison of the function duration of check_irq_vectors_for_cpu_disable() with/without the patch (32 online CPUs)
before patch:
[001] d..1 264.017803: check_irq_vectors_for_cpu_disable <-native_cpu_disable
[001] d..1 264.017886: clear_local_APIC <-native_cpu_disable
after patch:
[001] d..1 862.011847: check_irq_vectors_for_cpu_disable <-native_cpu_disable
[001] d..1 862.011850: clear_local_APIC <-native_cpu_disable
About (83 - 17) = 66 microsecond.

thanks,
Yu

Quoted 31 lines

P.

Quoted 29 lines
Cc Thomas Gleixner <[email protected]>
Cc Ingo Molnar <[email protected]>
Cc "H. Peter Anvin" <[email protected]>
Cc Prarit Bhargava <[email protected]>
Cc "Rafael J. Wysocki" <[email protected]>
Cc Len Brown <[email protected]>
---
 arch/x86/kernel/irq.c | 3 +++
 1 file changed, 3 insertions(+)
arch/x86/kernel/irq.c +3 -0
index 4d8183b..b9bdc1e 100644
@@ -394,6 +394,9 @@ int check_irq_vectors_for_cpu_disable(void)
 		    !cpumask_subset(&affinity_new, &online_new))
 			this_count++;
 	}
+	/* No need to check any further. */
+	if (!this_count)
+		return 0;
 
 	count = 0;
 	for_each_online_cpu(cpu) {
#3 lore raw
Quoted 10 lines · Chen Yu

I've just done a comparison of the function duration of check_irq_vectors_for_cpu_disable() with/without the patch (32 online CPUs)
before patch:
[001] d..1 264.017803: check_irq_vectors_for_cpu_disable <-native_cpu_disable
[001] d..1 264.017886: clear_local_APIC <-native_cpu_disable
after patch:
[001] d..1 862.011847: check_irq_vectors_for_cpu_disable <-native_cpu_disable
[001] d..1 862.011850: clear_local_APIC <-native_cpu_disable
About (83 - 17) = 66 microsecond.

Not really impressive if you subtract the tracer overhead. Though, if you want to micro optimize that, then you should do it proper and also break the for_each_cpu_loop() when the count reaches the required vectors.

Thanks,

tglx

arch/x86/kernel/irq.c +4 -2
@@ -411,8 +411,10 @@ int check_irq_vectors_for_cpu_disable(vo
 		for (vector = FIRST_EXTERNAL_VECTOR;
 		     vector < first_system_vector; vector++) {
 			if (!test_bit(vector, used_vectors) &&
-			    IS_ERR_OR_NULL(per_cpu(vector_irq, cpu)[vector]))
-			    count++;
+			    IS_ERR_OR_NULL(per_cpu(vector_irq, cpu)[vector])) {
+				if (++count == this_count)
+					return 0;
+			}
 		}
 	}