Re: [PATCH] sched/fair: Add null pointer check to pick_next_entity()

[Chris Mason]

On 4/16/25 10:19 AM, Rik van Riel wrote:
> On Wed, 2025-04-16 at 14:44 +0200, Peter Zijlstra wrote:
>> On Tue, Apr 15, 2025 at 10:02:35AM +0200, Peter Zijlstra wrote:
>>> On Mon, Apr 14, 2025 at 03:57:42PM -0400, Rik van Riel wrote:
>>>> On Wed, 2025-04-02 at 10:22 +0200, Peter Zijlstra wrote:
>>>>>
>>>>> Please confirm what the reason for overflow is.
>>>>>
>>>> Running a large enough sample size has its benefits.
>>>>
>>>> We have hit 3 out of the 4 warnings below.
>>>>
>>>> The only one we did not hit is the cfs_rq->avg_load != avg_load
>>>> warning.
>>>
>>> Fair enough, that one really isn't hard.
>>>
>>>> Most of the time we seem to hit the warnings from the
>>>> code that removes tasks from the runqueue, 
>>>
>>> *blink*..
>>
>> Which warning is getting hit on removal? The avg_vruntime mismatch?
>>
>> Also, which removal path? schedule()'s block path, or migration like?
> 
> The most common warning by far, hitting
> about 90% of the time we hit anything
> in avg_vruntime_validate is the
> WARN_ON_ONCE(cfs_rq->avg_vruntime != vruntime)
> 
> The most common code path to getting there,
> covering about 85% of the cases:
> 
> avg_vruntime_validate
> avg_vruntime_sub
> __dequeue_entity
> set_next_entity
> pick_task_fair
> pick_next_task_fair
> __pick_next_task
> pick_next_task
> __schedule
> schedule
> 
> 

When I was spot checking hosts, I only found these early in boot.  Rik,
did you find some that tripped later as well?

-chris