[PATCH] sched/fair: Correct CPU selection from isolated domain

[PATCH] sched/fair: Correct CPU selection from isolated domain

[wujing]

We encountered an issue where the kernel thread `ksmd` runs on the PMD
dedicated isolated core, leading to high latency in OVS packets.

Upon analysis, we discovered that this is caused by the current
select_idle_smt() function not taking the sched_domain mask into account.

Kernel version: linux-4.19.y

Signed-off-by: wujing <realwujing@qq.com>
Signed-off-by: QiLiang Yuan <yuanql9@chinatelecom.cn>
---
 kernel/sched/fair.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
index 09f82c84474b..0950cabfc1d0 100644
--- a/kernel/sched/fair.c
+++ b/kernel/sched/fair.c
@@ -6171,7 +6171,8 @@ static int select_idle_smt(struct task_struct *p, struct sched_domain *sd, int t
 		return -1;
 
 	for_each_cpu(cpu, cpu_smt_mask(target)) {
-		if (!cpumask_test_cpu(cpu, &p->cpus_allowed))
+		if (!cpumask_test_cpu(cpu, &p->cpus_allowed) ||
+			!cpumask_test_cpu(cpu, sched_domain_span(sd)))
 			continue;
 		if (available_idle_cpu(cpu))
 			return cpu;
-- 
2.45.2

[PATCH] sched/fair: Correct CPU selection from isolated domain

[wujing]

From: wujing <realwujing@gmail.com>

We encountered an issue where the kernel thread `ksmd` runs on the PMD
dedicated isolated core, leading to high latency in OVS packets.

Upon analysis, we discovered that this is caused by the current
select_idle_smt() function not taking the sched_domain mask into account.

Kernel version: linux-4.19.y

Cc: stable@vger.kernel.org # 4.19.x
Signed-off-by: wujing <realwujing@qq.com>
Signed-off-by: QiLiang Yuan <yuanql9@chinatelecom.cn>
---
 kernel/sched/fair.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
index 09f82c84474b..0950cabfc1d0 100644
--- a/kernel/sched/fair.c
+++ b/kernel/sched/fair.c
@@ -6171,7 +6171,8 @@ static int select_idle_smt(struct task_struct *p, struct sched_domain *sd, int t
 		return -1;
 
 	for_each_cpu(cpu, cpu_smt_mask(target)) {
-		if (!cpumask_test_cpu(cpu, &p->cpus_allowed))
+		if (!cpumask_test_cpu(cpu, &p->cpus_allowed) ||
+			!cpumask_test_cpu(cpu, sched_domain_span(sd)))
 			continue;
 		if (available_idle_cpu(cpu))
 			return cpu;
-- 
2.39.5

Re: [PATCH] sched/fair: Correct CPU selection from isolated domain

[Greg KH]

On Mon, Jan 06, 2025 at 05:26:34PM +0800, wujing wrote:
> From: wujing <realwujing@gmail.com>
> 
> We encountered an issue where the kernel thread `ksmd` runs on the PMD
> dedicated isolated core, leading to high latency in OVS packets.
> 
> Upon analysis, we discovered that this is caused by the current
> select_idle_smt() function not taking the sched_domain mask into account.
> 
> Kernel version: linux-4.19.y

You do know that 4.19.y is long end-of-life and totally insecure and
nothing that you should be using in anything, right?  Please move to a
more modern kernel version.

good luck!

greg k-h

[PATCH] sched/fair: Correct CPU selection from isolated domain

[wujing]

We encountered an issue where the kernel thread `ksmd` runs on the PMD
dedicated isolated core, leading to high latency in OVS packets.

Upon analysis, we discovered that this is caused by the current
select_idle_smt() function not taking the sched_domain mask into account.

Kernel version: linux-4.19.y

Signed-off-by: wujing <realwujing@qq.com>
Signed-off-by: QiLiang Yuan <yuanql9@chinatelecom.cn>
---
 kernel/sched/fair.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
index 09f82c84474b..0950cabfc1d0 100644
--- a/kernel/sched/fair.c
+++ b/kernel/sched/fair.c
@@ -6171,7 +6171,8 @@ static int select_idle_smt(struct task_struct *p, struct sched_domain *sd, int t
 		return -1;
 
 	for_each_cpu(cpu, cpu_smt_mask(target)) {
-		if (!cpumask_test_cpu(cpu, &p->cpus_allowed))
+		if (!cpumask_test_cpu(cpu, &p->cpus_allowed) ||
+			!cpumask_test_cpu(cpu, sched_domain_span(sd)))
 			continue;
 		if (available_idle_cpu(cpu))
 			return cpu;
-- 
2.45.2

Re: [PATCH] sched/fair: Correct CPU selection from isolated domain

[Peter Zijlstra]

On Tue, Jul 30, 2024 at 03:10:50PM +0800, wujing wrote:
> We encountered an issue where the kernel thread `ksmd` runs on the PMD
> dedicated isolated core, leading to high latency in OVS packets.
> 
> Upon analysis, we discovered that this is caused by the current
> select_idle_smt() function not taking the sched_domain mask into account.
> 
> Kernel version: linux-4.19.y

If you're trying to backport something, I think you forgot to Cc stable
and provide the proper upstream commit.

As is this isn't something I can do anything with. The patch does not
apply to any recent kernel and AFAICT this issue has long since been
fixed.

> 
> Signed-off-by: wujing <realwujing@qq.com>
> Signed-off-by: QiLiang Yuan <yuanql9@chinatelecom.cn>
> ---
>  kernel/sched/fair.c | 3 ++-
>  1 file changed, 2 insertions(+), 1 deletion(-)
> 
> diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> index 09f82c84474b..0950cabfc1d0 100644
> --- a/kernel/sched/fair.c
> +++ b/kernel/sched/fair.c
> @@ -6171,7 +6171,8 @@ static int select_idle_smt(struct task_struct *p, struct sched_domain *sd, int t
>  		return -1;
>  
>  	for_each_cpu(cpu, cpu_smt_mask(target)) {
> -		if (!cpumask_test_cpu(cpu, &p->cpus_allowed))
> +		if (!cpumask_test_cpu(cpu, &p->cpus_allowed) ||
> +			!cpumask_test_cpu(cpu, sched_domain_span(sd)))
>  			continue;
>  		if (available_idle_cpu(cpu))
>  			return cpu;
> -- 
> 2.45.2
> 

Re: [PATCH] sched/fair: Correct CPU selection from isolated domain

[zhengzucheng]

it has already been addressed by the following patches in tip/sched/core:

8aeaffef8c6e ("sched/fair: Take the scheduling domain into account in 
select_idle_smt()")
23d04d8c6b8e ("sched/fair: Take the scheduling domain into account in 
select_idle_core()")

在 2024/7/30 15:10, wujing 写道:
> We encountered an issue where the kernel thread `ksmd` runs on the PMD
> dedicated isolated core, leading to high latency in OVS packets.
>
> Upon analysis, we discovered that this is caused by the current
> select_idle_smt() function not taking the sched_domain mask into account.
>
> Kernel version: linux-4.19.y
>
> Signed-off-by: wujing <realwujing@qq.com>
> Signed-off-by: QiLiang Yuan <yuanql9@chinatelecom.cn>
> ---
>   kernel/sched/fair.c | 3 ++-
>   1 file changed, 2 insertions(+), 1 deletion(-)
>
> diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> index 09f82c84474b..0950cabfc1d0 100644
> --- a/kernel/sched/fair.c
> +++ b/kernel/sched/fair.c
> @@ -6171,7 +6171,8 @@ static int select_idle_smt(struct task_struct *p, struct sched_domain *sd, int t
>   		return -1;
>   
>   	for_each_cpu(cpu, cpu_smt_mask(target)) {
> -		if (!cpumask_test_cpu(cpu, &p->cpus_allowed))
> +		if (!cpumask_test_cpu(cpu, &p->cpus_allowed) ||
> +			!cpumask_test_cpu(cpu, sched_domain_span(sd)))
>   			continue;
>   		if (available_idle_cpu(cpu))
>   			return cpu;