Re: [PATCH v4 3/4] workqueue: Convert the idle_timer to a delayed_work

[Valentin Schneider <vschneid@redhat.com>]

On 31/10/22 08:49, Tejun Heo wrote:
> Hello,
>
> On Tue, Oct 04, 2022 at 04:05:20PM +0100, Valentin Schneider wrote:
>> +static void idle_reaper_fn(struct work_struct *work)
>>  {
>> -	struct worker_pool *pool = from_timer(pool, t, idle_timer);
>> +	struct delayed_work *dwork = to_delayed_work(work);
>> +	struct worker_pool *pool = container_of(dwork, struct worker_pool, idle_reaper_work);
>>
>>      raw_spin_lock_irq(&pool->lock);
>>
>>      while (too_many_workers(pool)) {
>>              struct worker *worker;
>>              unsigned long expires;
>> +		unsigned long now = jiffies;
>>
>> -		/* idle_list is kept in LIFO order, check the last one */
>> +		/* idle_list is kept in LIFO order, check the oldest entry */
>>              worker = list_entry(pool->idle_list.prev, struct worker, entry);
>>              expires = worker->last_active + IDLE_WORKER_TIMEOUT;
>>
>> -		if (time_before(jiffies, expires)) {
>> -			mod_timer(&pool->idle_timer, expires);
>
> So, one thing which bothers me is that the idle timer is supposed to go off
> spuriously periodically. The idea being that letting it expire spuriously
> should usually be cheaper than trying to update it constantly as workers
> wake up and sleep. Converting the timer to a delayed work makes spurious
> wakeups significantly more expensive tho as it's now a full scheduling
> event.
>

Right.

> Can we separate the timer and work item out so that we can kick off the work
> item iff there actually are tasks to kill?
>

One thing I can try to have a DIY delayed_work where the timer callback
doesn't just queue the work but first checks for too_many_workers(). This
will mostly likely result in a different behaviour as worker deletion will
then involve two pool->lock regions, but this will still catch long-idling
workers.

> Thanks.
>
> --
> tejun