Re: [PATCHSET] cgroup: introduce cgroup_taskset and consolidate subsys methods

[Frederic Weisbecker <fweisbec@gmail.com>]

On Wed, Aug 24, 2011 at 09:49:59AM +0200, Tejun Heo wrote:
> Hello, Frederic.
> 
> On Wed, Aug 24, 2011 at 03:14:30AM +0200, Frederic Weisbecker wrote:
> > >  0001-cgroup-subsys-attach_task-should-be-called-after-mig.patch
> > >  0002-cgroup-improve-old-cgroup-handling-in-cgroup_attach_.patch
> > >  0003-cgroup-introduce-cgroup_taskset-and-use-it-in-subsys.patch
> > >  0004-cgroup-don-t-use-subsys-can_attach_task-or-attach_ta.patch
> > >  0005-cgroup-cpuset-don-t-use-ss-pre_attach.patch
> > >  0006-cgroup-kill-subsys-can_attach_task-pre_attach-and-at.patch
> > 
> > I don't understand the point on patches 3,4,5,6
> > 
> > Why pushing the task iterations down to the subsystems?
> 
> I'll try again.
> 
> It seems like methods were added to serve the immediate need of the
> particular user at the time and that in turn led to addition of
> callbacks which were both superflous and incomplete (the bullet points
> in the original message list them).  This seems to have happened
> because extra interface was added without trying to make the existing
> interface complete.
> 
> The interface is complicated and cumbersome to use - are
> [can_]attach() called first or [can_]attach_task()?  What about
> cancelation?  What if a subsys wants to perform operations across
> multiple tasks atomically?
> 
> In general, iteration-by-callback is painful to use.  Establishing
> common context (be it synchronization domain or shared variables)
> becomes very cumbersome and implementation becomes fragmented and
> difficult to follow.  For example, imagine how it would be like to use
> list if we had call_for_each_list_entry(func, list_head) instead of
> the control-loop style iterators we have know.
> 
> So, using iterators enables making all relevant information to each
> stage of attach so that only one callback is required for each step -
> the way it should be.  In addition, it makes it far easier for
> subsystems to implement more involved logic in their methods.
> 
> I tried to make cgroup_freezer behave better which requires better
> synchronization against the freezer and, with the current interface,
> it's extremely ugly and painful.  The new interface is complete, easy
> to understand and use with far less subtleties.

Yeah it's true that the order between [can]attach/[can]attach_task plus
the added mess with pre_attach was not entirely sane. The fact we have
foo and foo_task is already a problem.

I guess we indeed need to sacrifice the iteration from the cgroup core
for that.