[PATCH] sched/cpuacct: fix percpu time accounting

[PATCH] sched/cpuacct: fix percpu time accounting

[Chengming Zhou]

The usage percpu data is the CPU time consumed on each CPU by all tasks
in this cgroup (including tasks lower in the hierarchy). When cpuacct_charge
called from other CPUs, we should use the CPU of task, not this CPU.

e.g.
    cpuacct_charge+1
    update_curr+332
    enqueue_entity+70
    enqueue_task_fair+169
    activate_task+57
    attach_task+46
    load_balance+1512
    run_rebalance_domains+451
    __do_softirq+282
    sysvec_apic_timer_interrupt+159
    asm_sysvec_apic_timer_interrupt+18
    native_safe_halt+11
    default_idle+10
    default_enter_idle+45
    cpuidle_enter_state+130
    cpuidle_enter+47
    do_idle+489
    cpu_startup_entry+25
    start_secondary+261
    secondary_startup_64_no_verify+176

Reported-by: Xuan Lu <luxuan.windniw@bytedance.com>
Reported-by: Minye Zhu <zhuminye@bytedance.com>
Co-developed-by: Qiang Wang <wangqiang.wq.frank@bytedance.com>
Signed-off-by: Qiang Wang <wangqiang.wq.frank@bytedance.com>
Signed-off-by: Chengming Zhou <zhouchengming@bytedance.com>
---
 kernel/sched/cpuacct.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/kernel/sched/cpuacct.c b/kernel/sched/cpuacct.c
index 893eece65bfd..aab51c88bd34 100644
--- a/kernel/sched/cpuacct.c
+++ b/kernel/sched/cpuacct.c
@@ -341,6 +341,7 @@ void cpuacct_charge(struct task_struct *tsk, u64 cputime)
 	struct cpuacct *ca;
 	int index = CPUACCT_STAT_SYSTEM;
 	struct pt_regs *regs = get_irq_regs() ? : task_pt_regs(tsk);
+	unsigned int cpu = task_cpu(tsk);
 
 	if (regs && user_mode(regs))
 		index = CPUACCT_STAT_USER;
@@ -348,7 +349,7 @@ void cpuacct_charge(struct task_struct *tsk, u64 cputime)
 	rcu_read_lock();
 
 	for (ca = task_ca(tsk); ca; ca = parent_ca(ca))
-		__this_cpu_add(ca->cpuusage->usages[index], cputime);
+		per_cpu_ptr(ca->cpuusage, cpu)->usages[index] += cputime;
 
 	rcu_read_unlock();
 }
-- 
2.11.0

Re: [PATCH] sched/cpuacct: fix percpu time accounting

[Chengming Zhou]

Ping guys. Any comments or suggestions?

On 2021/12/8 3:38 下午, Chengming Zhou wrote:
> The usage percpu data is the CPU time consumed on each CPU by all tasks
> in this cgroup (including tasks lower in the hierarchy). When cpuacct_charge
> called from other CPUs, we should use the CPU of task, not this CPU.
> 
> e.g.
>     cpuacct_charge+1
>     update_curr+332
>     enqueue_entity+70
>     enqueue_task_fair+169
>     activate_task+57
>     attach_task+46
>     load_balance+1512
>     run_rebalance_domains+451
>     __do_softirq+282
>     sysvec_apic_timer_interrupt+159
>     asm_sysvec_apic_timer_interrupt+18
>     native_safe_halt+11
>     default_idle+10
>     default_enter_idle+45
>     cpuidle_enter_state+130
>     cpuidle_enter+47
>     do_idle+489
>     cpu_startup_entry+25
>     start_secondary+261
>     secondary_startup_64_no_verify+176
> 
> Reported-by: Xuan Lu <luxuan.windniw@bytedance.com>
> Reported-by: Minye Zhu <zhuminye@bytedance.com>
> Co-developed-by: Qiang Wang <wangqiang.wq.frank@bytedance.com>
> Signed-off-by: Qiang Wang <wangqiang.wq.frank@bytedance.com>
> Signed-off-by: Chengming Zhou <zhouchengming@bytedance.com>
> ---
>  kernel/sched/cpuacct.c | 3 ++-
>  1 file changed, 2 insertions(+), 1 deletion(-)
> 
> diff --git a/kernel/sched/cpuacct.c b/kernel/sched/cpuacct.c
> index 893eece65bfd..aab51c88bd34 100644
> --- a/kernel/sched/cpuacct.c
> +++ b/kernel/sched/cpuacct.c
> @@ -341,6 +341,7 @@ void cpuacct_charge(struct task_struct *tsk, u64 cputime)
>  	struct cpuacct *ca;
>  	int index = CPUACCT_STAT_SYSTEM;
>  	struct pt_regs *regs = get_irq_regs() ? : task_pt_regs(tsk);
> +	unsigned int cpu = task_cpu(tsk);
>  
>  	if (regs && user_mode(regs))
>  		index = CPUACCT_STAT_USER;
> @@ -348,7 +349,7 @@ void cpuacct_charge(struct task_struct *tsk, u64 cputime)
>  	rcu_read_lock();
>  
>  	for (ca = task_ca(tsk); ca; ca = parent_ca(ca))
> -		__this_cpu_add(ca->cpuusage->usages[index], cputime);
> +		per_cpu_ptr(ca->cpuusage, cpu)->usages[index] += cputime;
>  
>  	rcu_read_unlock();
>  }