Re: [RFC PATCH 3/3] mm/memcg: Allow the task_obj optimization only on non-PREEMPTIBLE kernels.

[Waiman Long <longman@redhat.com>]

On 1/3/22 09:44, Sebastian Andrzej Siewior wrote:
> On 2021-12-23 16:48:41 [-0500], Waiman Long wrote:
>> On 12/22/21 06:41, Sebastian Andrzej Siewior wrote:
>>> Based on my understanding the optimisation with task_obj for in_task()
>>> mask sense on non-PREEMPTIBLE kernels because preempt_disable()/enable()
>>> is optimized away. This could be then restricted to !CONFIG_PREEMPTION kernel
>>> instead to only PREEMPT_RT.
>>> With CONFIG_PREEMPT_DYNAMIC a non-PREEMPTIBLE kernel can also be
>>> configured but these kernels always have preempt_disable()/enable()
>>> present so it probably makes no sense here for the optimisation.
>>>
>>> Restrict the optimisation to !CONFIG_PREEMPTION kernels.
>>>
>>> Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
>> If PREEMPT_DYNAMIC is selected, PREEMPTION will also be set. My
>> understanding is that some distros are going to use PREEMPT_DYNAMIC, but
>> default to PREEMPT_VOLUNTARY. So I don't believe it is a good idea to
>> disable the optimization based on PREEMPTION alone.
> So there is a benefit to this even if preempt_disable() is not optimized
> away? My understanding was that this depends on preempt_disable() being
> optimized away.
> Is there something you recommend as a benchmark where I could get some
> numbers?

In the case of PREEMPT_DYNAMIC, it depends on the default setting which 
is used by most users. I will support disabling the optimization if 
defined(CONFIG_PREEMPT_RT) || defined(CONFIG_PREEMPT), just not by 
CONFIG_)PREEMPTION alone.

As for microbenchmark, something that makes a lot of calls to malloc() 
or related allocations can be used.

Cheers,
Longman