[v2] sched/core: Adapt WARN_DOUBLE_CLOCK machinery for core-sched

  When sched_core_enabled(), we sometimes need to call update_rq_clock()
to update the rq clock of sibling CPUs on the same core, before that we
need to clear RQCF_UPDATED of rq->clock_update_flags to avoid the
WARN_DOUBLE_CLOCK warning. Because at this time the rq->clock_update_flags
of sibling CPUs may be RQCF_UPDATED. If sched_core_enabled(), we will get
a core wide rq->lock, so at this point we can safely clear RQCF_UPDATED of
rq->clock_update_flags of all CPUs on this core to avoid the
WARN_DOUBLE_CLOCK warning.

We cannot clear rq->clock_update_flags of other cpus on the same core in
rq_pin_lock(). Because in some functions, we will temporarily give up
core wide rq->lock, and then use raw_spin_rq_lock() to obtain core wide
rq->lock, such as newidle_balance() and _double_lock_balance().

Steps to reproduce:
1. Enable CONFIG_SCHED_DEBUG and CONFIG_SCHED_CORE when compiling
   the kernel
2. echo 1 > /sys/kernel/debug/clear_warn_once
   echo "WARN_DOUBLE_CLOCK" > /sys/kernel/debug/sched/features
3. Run the linux/tools/testing/selftests/sched/cs_prctl_test test

Signed-off-by: Hao Jia <jiahao.os@bytedance.com>
---
 - Adapt WARN_DOUBLE_CLOCK machinery for core-sched instead of clearing
   WARN_DOUBLE_CLOCK warning one by one.
 - Modify commit information
 [v1] https://lore.kernel.org/all/20221206070550.31763-1-jiahao.os@bytedance.com

 kernel/sched/core.c | 17 +++++++++++++++++
 1 file changed, 17 insertions(+)
  

On Wed, Feb 15, 2023 at 03:39:27PM +0800, Hao Jia wrote:
> When sched_core_enabled(), we sometimes need to call update_rq_clock()
> to update the rq clock of sibling CPUs on the same core, before that we
> need to clear RQCF_UPDATED of rq->clock_update_flags to avoid the
> WARN_DOUBLE_CLOCK warning. Because at this time the rq->clock_update_flags
> of sibling CPUs may be RQCF_UPDATED. If sched_core_enabled(), we will get
> a core wide rq->lock, so at this point we can safely clear RQCF_UPDATED of
> rq->clock_update_flags of all CPUs on this core to avoid the
> WARN_DOUBLE_CLOCK warning.
> 
> We cannot clear rq->clock_update_flags of other cpus on the same core in
> rq_pin_lock(). Because in some functions, we will temporarily give up
> core wide rq->lock, and then use raw_spin_rq_lock() to obtain core wide
> rq->lock, such as newidle_balance() and _double_lock_balance().
> 
> Steps to reproduce:
> 1. Enable CONFIG_SCHED_DEBUG and CONFIG_SCHED_CORE when compiling
>    the kernel
> 2. echo 1 > /sys/kernel/debug/clear_warn_once
>    echo "WARN_DOUBLE_CLOCK" > /sys/kernel/debug/sched/features
> 3. Run the linux/tools/testing/selftests/sched/cs_prctl_test test
> 
> Signed-off-by: Hao Jia <jiahao.os@bytedance.com>
> ---
>  - Adapt WARN_DOUBLE_CLOCK machinery for core-sched instead of clearing
>    WARN_DOUBLE_CLOCK warning one by one.
>  - Modify commit information
>  [v1] https://lore.kernel.org/all/20221206070550.31763-1-jiahao.os@bytedance.com
> 
>  kernel/sched/core.c | 17 +++++++++++++++++
>  1 file changed, 17 insertions(+)
> 
> diff --git a/kernel/sched/core.c b/kernel/sched/core.c
> index e838feb6adc5..16a33e5adb77 100644
> --- a/kernel/sched/core.c
> +++ b/kernel/sched/core.c
> @@ -427,11 +427,27 @@ void sched_core_put(void)
>  		schedule_work(&_work);
>  }
>  
> +static inline void sched_core_rq_clock_clear_update(struct rq *rq)
> +{
> +#ifdef CONFIG_SCHED_DEBUG
> +	const struct cpumask *smt_mask;
> +	int i;
> +
> +	if (rq->core_enabled) {
> +		smt_mask = cpu_smt_mask(rq->cpu);
> +		for_each_cpu(i, smt_mask) {
> +			if (rq->cpu != i)
> +				cpu_rq(i)->clock_update_flags &= (RQCF_REQ_SKIP|RQCF_ACT_SKIP);
> +		}
> +	}
> +#endif

So sort of ok, but that function name.... so long :/

> +}
>  #else /* !CONFIG_SCHED_CORE */
>  
>  static inline void sched_core_enqueue(struct rq *rq, struct task_struct *p) { }
>  static inline void
>  sched_core_dequeue(struct rq *rq, struct task_struct *p, int flags) { }
> +static inline void sched_core_rq_clock_clear_update(struct rq *rq) { }
>  
>  #endif /* CONFIG_SCHED_CORE */
>  
> @@ -546,6 +562,7 @@ void raw_spin_rq_lock_nested(struct rq *rq, int subclass)
>  		if (likely(lock == __rq_lockp(rq))) {
>  			/* preempt_count *MUST* be > 1 */
>  			preempt_enable_no_resched();
> +			sched_core_rq_clock_clear_update(rq);
>  			return;
>  		}
>  		raw_spin_unlock(lock);

This otoh don't make much sense. Why put it here and not extend
rq_pin_lock()?

That is, what's wrong with something like so?

---

diff --git a/kernel/sched/sched.h b/kernel/sched/sched.h
index 771f8ddb7053..c1a92eced930 100644
--- a/kernel/sched/sched.h
+++ b/kernel/sched/sched.h
@@ -1571,11 +1571,18 @@ static inline void rq_pin_lock(struct rq *rq, struct rq_flags *rf)
 	rf->cookie = lockdep_pin_lock(__rq_lockp(rq));
 
 #ifdef CONFIG_SCHED_DEBUG
-	rq->clock_update_flags &= (RQCF_REQ_SKIP|RQCF_ACT_SKIP);
-	rf->clock_update_flags = 0;
 #ifdef CONFIG_SMP
 	SCHED_WARN_ON(rq->balance_callback && rq->balance_callback != &balance_push_callback);
 #endif
+	rf->clock_update_flags = 0;
+	if (sched_core_enabled()) {
+		int i;
+
+		for_each_cpu(i, cpu_smt_mask(rq->cpu))
+			cpu_rq(i)->clock_update_flags &= (RQCF_REQ_SKIP|RQCF_ACT_SKIP);
+	} else {
+		rq->clock_update_flags &= (RQCF_REQ_SKIP|RQCF_ACT_SKIP);
+	}
 #endif
 }
  

On 2023/2/20 Peter Zijlstra wrote:
> On Wed, Feb 15, 2023 at 03:39:27PM +0800, Hao Jia wrote:
>> When sched_core_enabled(), we sometimes need to call update_rq_clock()
>> to update the rq clock of sibling CPUs on the same core, before that we
>> need to clear RQCF_UPDATED of rq->clock_update_flags to avoid the
>> WARN_DOUBLE_CLOCK warning. Because at this time the rq->clock_update_flags
>> of sibling CPUs may be RQCF_UPDATED. If sched_core_enabled(), we will get
>> a core wide rq->lock, so at this point we can safely clear RQCF_UPDATED of
>> rq->clock_update_flags of all CPUs on this core to avoid the
>> WARN_DOUBLE_CLOCK warning.
>>
>> We cannot clear rq->clock_update_flags of other cpus on the same core in
>> rq_pin_lock(). Because in some functions, we will temporarily give up
>> core wide rq->lock, and then use raw_spin_rq_lock() to obtain core wide
>> rq->lock, such as newidle_balance() and _double_lock_balance().
>>
>> Steps to reproduce:
>> 1. Enable CONFIG_SCHED_DEBUG and CONFIG_SCHED_CORE when compiling
>>     the kernel
>> 2. echo 1 > /sys/kernel/debug/clear_warn_once
>>     echo "WARN_DOUBLE_CLOCK" > /sys/kernel/debug/sched/features
>> 3. Run the linux/tools/testing/selftests/sched/cs_prctl_test test
>>
>> Signed-off-by: Hao Jia <jiahao.os@bytedance.com>
>> ---
>>   - Adapt WARN_DOUBLE_CLOCK machinery for core-sched instead of clearing
>>     WARN_DOUBLE_CLOCK warning one by one.
>>   - Modify commit information
>>   [v1] https://lore.kernel.org/all/20221206070550.31763-1-jiahao.os@bytedance.com
>>
>>   kernel/sched/core.c | 17 +++++++++++++++++
>>   1 file changed, 17 insertions(+)
>>
>> diff --git a/kernel/sched/core.c b/kernel/sched/core.c
>> index e838feb6adc5..16a33e5adb77 100644
>> --- a/kernel/sched/core.c
>> +++ b/kernel/sched/core.c
>> @@ -427,11 +427,27 @@ void sched_core_put(void)
>>   		schedule_work(&_work);
>>   }
>>   
>> +static inline void sched_core_rq_clock_clear_update(struct rq *rq)
>> +{
>> +#ifdef CONFIG_SCHED_DEBUG
>> +	const struct cpumask *smt_mask;
>> +	int i;
>> +
>> +	if (rq->core_enabled) {
>> +		smt_mask = cpu_smt_mask(rq->cpu);
>> +		for_each_cpu(i, smt_mask) {
>> +			if (rq->cpu != i)
>> +				cpu_rq(i)->clock_update_flags &= (RQCF_REQ_SKIP|RQCF_ACT_SKIP);
>> +		}
>> +	}
>> +#endif
> 
> So sort of ok, but that function name.... so long :/

What about clear_core_clock_updated()?
Indeed, a good function name is difficult.

Thanks,
Hao
> 
>> +}
>>   #else /* !CONFIG_SCHED_CORE */
>>   
>>   static inline void sched_core_enqueue(struct rq *rq, struct task_struct *p) { }
>>   static inline void
>>   sched_core_dequeue(struct rq *rq, struct task_struct *p, int flags) { }
>> +static inline void sched_core_rq_clock_clear_update(struct rq *rq) { }
>>   
>>   #endif /* CONFIG_SCHED_CORE */
>>   
>> @@ -546,6 +562,7 @@ void raw_spin_rq_lock_nested(struct rq *rq, int subclass)
>>   		if (likely(lock == __rq_lockp(rq))) {
>>   			/* preempt_count *MUST* be > 1 */
>>   			preempt_enable_no_resched();
>> +			sched_core_rq_clock_clear_update(rq);
>>   			return;
>>   		}
>>   		raw_spin_unlock(lock);
> 
> This otoh don't make much sense. Why put it here and not extend
> rq_pin_lock()?
> 
> That is, what's wrong with something like so?

We sometimes use rq_pin_lock() and raw_spin_rq_lock() separately.
in some functions（）, we will temporarily give up
core wide rq->lock, and then use raw_spin_rq_lock() to obtain core wide
rq->lock, at this time the rq->clock_update_flags may be RQCF_UPDATED.
such as newidle_balance() and _double_lock_balance().


Perhaps it is more flexible to use rq_pin_lock() and raw_spin_rq_lock() 
separately? such as:

newidle_balance()
     rq_unpin_lock()
     raw_spin_rq_unlock()

     update_blocked_averages()
         update_rq_clock(rq)    <-- RQCF_UPDATED

     raw_spin_rq_lock(); （1）
     rq_repin_lock()           <-- we need *restore* to RQCF_UPDATED


We cannot replace raw_spin_rq_lock() in (1) with rq_lock(), because this 
will set rf->clock_update_flags to 0,
This may cause assert_clock_updated() to be triggered later.

Thinks,
Hao
> 
> ---
> 
> diff --git a/kernel/sched/sched.h b/kernel/sched/sched.h
> index 771f8ddb7053..c1a92eced930 100644
> --- a/kernel/sched/sched.h
> +++ b/kernel/sched/sched.h
> @@ -1571,11 +1571,18 @@ static inline void rq_pin_lock(struct rq *rq, struct rq_flags *rf)
>   	rf->cookie = lockdep_pin_lock(__rq_lockp(rq));
>   
>   #ifdef CONFIG_SCHED_DEBUG
> -	rq->clock_update_flags &= (RQCF_REQ_SKIP|RQCF_ACT_SKIP);
> -	rf->clock_update_flags = 0;
>   #ifdef CONFIG_SMP
>   	SCHED_WARN_ON(rq->balance_callback && rq->balance_callback != &balance_push_callback);
>   #endif
> +	rf->clock_update_flags = 0;
> +	if (sched_core_enabled()) {
> +		int i;
> +
> +		for_each_cpu(i, cpu_smt_mask(rq->cpu))
> +			cpu_rq(i)->clock_update_flags &= (RQCF_REQ_SKIP|RQCF_ACT_SKIP);
> +	} else {
> +		rq->clock_update_flags &= (RQCF_REQ_SKIP|RQCF_ACT_SKIP);
> +	}
>   #endif
>   }
>