Re: [PATCH 1/3] fork: dynamically allocate cache array for vmapped stacks using cpuhp

[Hoeun Ryu]

On Sun, Feb 5, 2017 at 7:18 PM, Michal Hocko <mhocko@xxxxxxxxxx> wrote:
> On Sat 04-02-17 11:01:32, Hoeun Ryu wrote:
>> On Sat, Feb 4, 2017 at 2:52 AM, Andy Lutomirski <luto@xxxxxxxxxxxxxx> wrote:
>> > On Fri, Feb 3, 2017 at 8:42 AM, Hoeun Ryu <hoeun.ryu@xxxxxxxxx> wrote:
>> >> On Sat, Feb 4, 2017 at 12:39 AM, Michal Hocko <mhocko@xxxxxxxxxx> wrote:
>> >>> On Sat 04-02-17 00:30:05, Hoeun Ryu wrote:
>> >>>>  Using virtually mapped stack, kernel stacks are allocated via vmalloc.
>> >>>> In the current implementation, two stacks per cpu can be cached when
>> >>>> tasks are freed and the cached stacks are used again in task duplications.
>> >>>> but the array for the cached stacks is statically allocated by per-cpu api.
>> >>>>  In this new implementation, the array for the cached stacks are dynamically
>> >>>> allocted and freed by cpu hotplug callbacks and the cached stacks are freed
>> >>>> when cpu is down. setup for cpu hotplug is established in fork_init().
>> >>>
>> >>> Why do we want this? I can see that the follow up patch makes the number
>> >>> configurable but the changelog doesn't describe the motivation for that.
>> >>> Which workload would benefit from a higher value?
>> >>>
>> >>
>> >> The key difference of this implementation, the cached stacks for a cpu
>> >> is freed when a cpu is down.
>> >> so the cached stacks are no longer wasted.
>> >> In the current implementation, the cached stacks for a cpu still
>> >> remain on the system when a cpu is down.
>> >> I think we could imagine what if a machine has many cpus and someone
>> >> wants to have bigger size of stack caches.
>> >
>> > Then how about just registering a simple hotplug hook to free the
>> > stacks without worrying about freeing the tiny array as well?
>> >
>>
>> Michal, What do you think about it. it sounds fair enough.
>
> This is what I've tried to suggest in the other reply.

OK, I'll work on patch1/2 again and drop patch3.

> --
> Michal Hocko
> SUSE Labs