Re: [RFD] Task counter: cgroup core feature or cgroup subsystem? (was Re: [PATCH 0/8 v3] cgroups: Task counter subsystem)

[Li Zefan <lizf@cn.fujitsu.com>]

Frederic Weisbecker wrote:
> On Tue, Aug 23, 2011 at 09:07:59AM -0700, Paul Menage wrote:
>> On Thu, Aug 18, 2011 at 7:33 AM, Frederic Weisbecker <fweisbec@gmail.com> wrote:
>>>
>>> So the problem with the task counter as a subsystem is that you could
>>> mount it in your systemd cgroups hierarchy but then it's not anymore
>>> available for those who want to use containers.
>>
>> Another possible option is something that I prototyped a couple of
>> years ago, but dropped due to lack of compelling need and demand - the
>> ability to have subsystems that can be bound on multiple subsystems at
>> once. See
>>
>> http://lkml.indiana.edu/hypermail/linux/kernel/0907.0/00574.html
>> http://lkml.indiana.edu/hypermail/linux/kernel/0907.0/00576.html
>> http://lkml.indiana.edu/hypermail/linux/kernel/0907.0/00577.html
>>
>> It's applicable to subsystems whose state isn't tied to any specific
>> single resource in the kernel outside of cgroups (so e.g. the CPU
>> scheduler couldn't be usefully multi-bindable, since the CPU cgroup
>> state is tied to the machine's single CPU scheduler).
>>
>> In the end I didn't work further on it, since it seemed that most
>> things that needed to be available to multiple hierarchies could more
>> simply be added to the core cgroups subsystem and automatically be
>> available on all hierarchies. But the point about tracking overhead
>> for fork/exit is certainly something that could make this worthwhile.
> 
> That sounds like a perfect fit. I like that much better because there
> should be no noticeable overhead when the task counter subsys is
> nowhere mounted, compared to a pure core feature.
> 

I had a patchset that can make things even better. The patchset is
to bind/unbind a subsystem to/from an existing cgroup hierarchy.

So if you found you need to use task couter but it's not avaiable
in the hierarchy you've set up, you can do this:

	# mount -o remount,task_num xxx /cgroup

Currently the above operation is supported only if there's no child
cgroups in /cgroup.

> So I'm going to continue to work on that task counter subsystem and
> I will unearth your old patch afterward to make that work on several
> mountpoints once we are sure this is needed for systemd.
> 
> It seems your patch doesn't handle the ->fork() and ->exit() calls.
> We probably need a quick access to states of multi-subsystems from
> the task, some lists available from task->cgroups, I don't know yet.
>