Re: [PATCH] sched/fair: Distribute burst-woken tasks when rq depth reaches 3x average

[Li kunyu]

Hi Kayra,

On 2026/11/5 2:10, Kayra Cizmeci wrote:
> 
> Hi!
> 
> On Wed, Nov 04, 2026 at 05:18:42PM +0800, Li kunyu wrote:
> 
> FWIW, This doesn't apply to the mainline:
> 
> Patch failed at 0001 sched/fair: Distribute burst-woken tasks when rq depth reaches 3x average
> error: patch failed: kernel/sched/fair.c:9972
> error: kernel/sched/fair.c: patch does not apply

Today, I updated the warehouse branch to the latest version, addressed the
identified issues, and performed another thorough test.

>> When a task burst-wakes a large number of tasks in a short period, tasks
>> can pile up on a single CPU regardless of whether the system has idle
>> CPUs or NO_HZ enabled.
>>
>> Track the average runqueue depth (scaled by SCHED_CAPACITY_SCALE to
>> preserve fractional precision) in root_domain during the periodic
>> load-balance pass.
>>
>> In select_task_rq_fair(), when the target runqueue depth reaches 3x the
>> system-wide average, fall back to cpumask_any_and_distribute(p->cpus_ptr,
>> cpu_active_mask) to rotate waking tasks across allowed CPUs and reset the
>> periodic load balance intervals so regular load balancing can promptly
>> smooth out the distribution.
>>
>> Suggested-by: Peter Zijlstra <peterz@xxxxxxxxxxxxx>
>> Link: https://lore.kernel.org/all/20261021125419.GK2823843@xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx/
>> Signed-off-by: Li kunyu <likunyu@xxxxxxxxxxxxxx>
>> ---
>>  kernel/sched/fair.c     | 70 ++++++++++++++++++++++++++++++++++++++---
>>  kernel/sched/sched.h    |  3 ++
>>  kernel/sched/topology.c |  2 ++
>>  3 files changed, 71 insertions(+), 4 deletions(-)
>>
>> diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
>> index 6e005f1e8dc9..e4d278ac46c4 100644
>> --- a/kernel/sched/fair.c
>> +++ b/kernel/sched/fair.c
>> @@ -9903,6 +9903,38 @@ static int find_energy_efficient_cpu(struct task_struct *p, int prev_cpu)
>>  	return target;
>>  }
>>  
>> +/*
>> + * Check if a CPU's rq depth reaches 3x the system-wide average tracked
>> + * in root_domain (scaled by SCHED_CAPACITY_SCALE, floored at 1 task).
>> + */
>> +static inline bool rq_burst_overloaded(int cpu)
>> +{
>> +	struct rq *rq = cpu_rq(cpu);
>> +	unsigned long avg;
>> +
>> +	if (unlikely(!rq->rd))
>> +		return false;
>> +
>> +	avg = max_t(unsigned long, READ_ONCE(rq->rd->avg_nr_running),
>> +		    SCHED_CAPACITY_SCALE);
>> +
>> +	return (READ_ONCE(rq->nr_running) << SCHED_CAPACITY_SHIFT) >= avg * 3;
>> +}
>> +
>> +static inline void reset_periodic_lb_interval(int cpu)
>> +{
>> +	struct sched_domain *sd;
>> +
>> +	/*
>> +	 * Only the most basic values are updated here to avoid excessive
>> +	 * cost.
>> +	 */
>> +	for_each_domain(cpu, sd) {
>> +		if (READ_ONCE(sd->balance_interval) > sd->min_interval)
>> +			WRITE_ONCE(sd->balance_interval, sd->min_interval);
>> +	}
>> +}
>> +
>>  /*
>>   * select_task_rq_fair: Select target runqueue for the waking task in domains
>>   * that have the relevant SD flag set. In practice, this is SD_BALANCE_WAKE,
>> @@ -9972,11 +10004,19 @@ select_task_rq_fair(struct task_struct *p, int prev_cpu, int wake_flags)
>>  
>>  	/* Slow path */
>>  	if (unlikely(sd))
>> -		return sched_balance_find_dst_cpu(sd, p, cpu, prev_cpu, sd_flag);
>> -
>> +		new_cpu = sched_balance_find_dst_cpu(sd, p, cpu, prev_cpu, sd_flag);
>>  	/* Fast path */
>> -	if (wake_flags & WF_TTWU)
>> -		return select_idle_sibling(p, prev_cpu, new_cpu);
>> +	else if (wake_flags & WF_TTWU)
>> +		new_cpu = select_idle_sibling(p, prev_cpu, new_cpu);
>> +
>> +	if (unlikely(rq_burst_overloaded(new_cpu))) {
>> +		int dist_cpu = cpumask_any_and_distribute(p->cpus_ptr, cpu_active_mask);
>> +
>> +		if (likely(dist_cpu < nr_cpumask_bits)) {
>> +			reset_periodic_lb_interval(new_cpu);
> 
> Why are we calling this for new_cpu?

Here, the existing logic is reused, with new_cpu used as the final return value.
It remains unclear whether this approach is appropriate. If it causes no
practical impact, it seems reasonable to keep it unchanged.

>> +			return dist_cpu;
>> +		}
>> +	}
>>  
>>  	return new_cpu;
>>  }
>> @@ -11685,6 +11725,7 @@ struct sg_lb_stats {
>>  struct sd_lb_stats {
>>  	struct sched_group *busiest;		/* Busiest group in this sd */
>>  	struct sched_group *local;		/* Local group in this sd */
>> +	unsigned long total_nr_running;		/* Sum of nr_running in this sd */
>>  	unsigned long total_load;		/* Total load of all groups in sd */
>>  	unsigned long total_capacity;		/* Total capacity of all groups in sd */
>>  	unsigned long avg_load;			/* Average load across all groups in sd */
>> @@ -12964,6 +13005,7 @@ static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sd
>>  		/* Now, start updating sd_lb_stats */
>>  		sds->total_load += sgs->group_load;
>>  		sds->total_capacity += sgs->group_capacity;
>> +		sds->total_nr_running += sgs->sum_nr_running;
>>  
>>  		sum_util += sgs->group_util;
>>  		sg = sg->next;
>> @@ -12982,11 +13024,31 @@ static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sd
>>  		env->fbq_type = fbq_classify_group(&sds->busiest_stat);
>>  
>>  	if (!env->sd->parent) {
>> +		unsigned int weight = env->sd->span_weight;
> 
> Uh-oh. There's something wrong, no?
> 
> span_weight equals all of the CPU's in the domain.
> So in sched_balance_rq() we call sched_balance_find_src_group() and that
> calls update_sd_lb_stats().
> 
> But it just retries balancing when all of the tasks
> in the rq is pinned. In that case the CPU is cleared from the env->cpus
> bitmap. But sum_nr_running is calculated for all env->cpus while
> span_weight is for all of the CPU's in that domain.
> 
> BTW, in theory this could happen multiple times in a row and
> it could make things even worse.
> 
> 
> Or am I missing something?

I have also noticed this. Let us optimize it based on the number of env->cpus.

>> +
>>  		/* update overload indicator if we are at root domain */
>>  		set_rd_overloaded(env->dst_rq->rd, sg_overloaded);
>>  
>>  		/* Update over-utilization (tipping point, U >= 0) indicator */
>>  		set_rd_overutilized(env->dst_rq->rd, sg_overutilized);
>> +
>> +		/*
>> +		 * We update the average task count at the top-level (root)
>> +		 * domain, as this incurs the least overhead and comes
>> +		 * closest to the ideal scenario. The trade-off is that we
>> +		 * must accept some temporary data inaccuracy.
>> +		 */
>> +		if (weight) {
>> +			unsigned long cur_avg;
>> +			unsigned long old_avg;
>> +
>> +			cur_avg = (sds->total_nr_running << SCHED_CAPACITY_SHIFT) / weight;
>> +			old_avg = READ_ONCE(env->dst_rq->rd->avg_nr_running);
>> +			if (old_avg)
>> +				cur_avg = (old_avg * 7 + cur_avg) >> 3;
>> +
>> +			WRITE_ONCE(env->dst_rq->rd->avg_nr_running, cur_avg);
>> +		}
>>  	} else if (sg_overutilized) {
>>  		set_rd_overutilized(env->dst_rq->rd, sg_overutilized);
>>  	}
>> diff --git a/kernel/sched/sched.h b/kernel/sched/sched.h
>> index e656c7059bf8..4147e043407c 100644
>> --- a/kernel/sched/sched.h
>> +++ b/kernel/sched/sched.h
>> @@ -1077,6 +1077,9 @@ struct root_domain {
>>  	cpumask_var_t		rto_mask;
>>  	struct cpupri		cpupri;
>>  
>> +	/* Scaled (SCHED_CAPACITY_SCALE) average per-CPU rq depth */
>> +	unsigned long		avg_nr_running;
>> +
>>  	/*
>>  	 * NULL-terminated list of performance domains intersecting with the
>>  	 * CPUs of the rd. Protected by RCU.
>> diff --git a/kernel/sched/topology.c b/kernel/sched/topology.c
>> index 3dab0253976f..41c3b9e30d65 100644
>> --- a/kernel/sched/topology.c
>> +++ b/kernel/sched/topology.c
>> @@ -558,6 +558,8 @@ static int init_rootdomain(struct root_domain *rd)
>>  
>>  	if (cpupri_init(&rd->cpupri) != 0)
>>  		goto free_cpudl;
>> +
>> +	rd->avg_nr_running = SCHED_CAPACITY_SCALE;
>>  	return 0;
>>  
>>  free_cpudl:
>> -- 
>> 2.47.3
>>
> 
> BTW, if we are looking to nr_running for this how about atleast looking
> at our dist_cpu and not returning it if its nr_running is bigger
> than the other CPU. Right now we are looking only at the average.
> 
> Would this work for you? Peter/Li?

I added this condition to the new patch to determine whether it improves the
test results or reduces the impact. Ideally, if the effect proves satisfactory,
this condition should be included.

> Thanks,
> 
> Kayra
> 

In the next round of testing, I will also follow up on the other suggestions to
ensure that the patch meets the quality bar.

-- 
thanks,
kunyu