Re: [PATCH] sched/fair: Rate limit calls to update_blocked_averages() for NOHZ

[Tim Chen <tim.c.chen@linux.intel.com>]

On 5/11/21 8:25 AM, Vincent Guittot wrote:
> Hi Tim,
> 
> Sometimes, we want to set this_rq->next_balance backward compared to
> its current value. When a CPU is busy, the balance interval is
> multiplied by busy_factor which is set to 16 by default. On SMT2 sched
> domain level, it means that the interval will be 32ms when busy
> instead of 2ms. But if a busy load balance happens just before
> becoming idle, the this_rq->next_balance will be set 32ms later
> whereas it should go down to 2ms as the CPU is now idle. And this
> becomes even worse when you have 128 CPUs at die sched_domain level
> because the idle CPU will have to wait 2048ms instead of the correct
> 128ms interval.
> 
>>
>>  out:
>>         /* Move the next balance forward */
>> -       if (time_after(this_rq->next_balance, next_balance))
>> +       if (time_after(next_balance, this_rq->next_balance))
> 
> The current comparison is correct but next_balance should not be in the past.

I understand then the intention is after the update,
this_rq->next_balance should have a minimum value of jiffies+1. So
we will need

 out:
        /* Move the next balance forward */
+       this_rq->next_balance = max(jiffies+1, this_rq->next_balance);
        if (time_after(this_rq->next_balance, next_balance))
                this_rq->next_balance = next_balance;

as next_balance computed will be at least jiffies+1 after your fix to
update_next_balance(). We still need to take care of the case when 
this_rq->next_balance <= jiffies.

So combining with your suggestion on the fix to update_next_balance(),
the fix will be

diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
index 1d75af1ecfb4..2dc471c1511c 100644
--- a/kernel/sched/fair.c
+++ b/kernel/sched/fair.c
@@ -9901,7 +9901,7 @@ update_next_balance(struct sched_domain *sd, unsigned long *next_balance)
 
        /* used by idle balance, so cpu_busy = 0 */
        interval = get_sd_balance_interval(sd, 0);
-       next = sd->last_balance + interval;
+       next = max(jiffies+1, sd->last_balance + interval);
 
        if (time_after(*next_balance, next))
                *next_balance = next;
@@ -10681,6 +10681,7 @@ static int newidle_balance(struct rq *this_rq, struct rq_flags *rf)
 
 out:
        /* Move the next balance forward */
+       this_rq->next_balance = max(jiffies+1, this_rq->next_balance);
        if (time_after(this_rq->next_balance, next_balance))
                this_rq->next_balance = next_balance;


> 
> update_next_balance() is only used in newidle_balance() so we could
> modify it to  have:
> 
> next = max(jiffies+1, next = sd->last_balance + interval)

Is the extra assignment "next = sd->last_balance + interval" needed?
This seems more straight forward:

next = max(jiffies+1, sd->last_balance + interval)

I will try to get the benchmark folks to do another run with this change.  
Hopefully I'll get some bandwidth from them soon.

Thanks.

Tim