Re: [PATCH 5/7] cgroups: Ability to stop res charge propagation on bounded ancestor

[Frederic Weisbecker <fweisbec@gmail.com>]

On Tue, Jul 12, 2011 at 09:11:31AM +0900, KAMEZAWA Hiroyuki wrote:
> On Mon, 11 Jul 2011 16:15:04 +0200
> Frederic Weisbecker <fweisbec@gmail.com> wrote:
> 
> > Moving a task from a cgroup to another may require to substract
> > its resource charge from the old cgroup and add it to the new one.
> > 
> > For this to happen, the uncharge/charge propagation can just stop
> > when we reach the common ancestor for the two cgroups. Further
> > the performance reasons, we also want to avoid to temporarily
> > overload the common ancestors with a non-accurate resource
> > counter usage if we charge first the new cgroup and uncharge the
> > old one thereafter. This is going to be a requirement for the coming
> > max number of task subsystem.
> > 
> > To solve this, provide a pair of new API that can charge/uncharge
> > a resource counter until we reach a given ancestor.
> > 
> > Signed-off-by: Frederic Weisbecker <fweisbec@gmail.com>
> > Cc: Paul Menage <menage@google.com>
> > Cc: Li Zefan <lizf@cn.fujitsu.com>
> > Cc: Johannes Weiner <hannes@cmpxchg.org>
> > Cc: Aditya Kali <adityakali@google.com>
> 
> 
> Hmm, do you have the number to show the benefit of this new function ?
> And....tasks is moving among cgroups so frequently as to show the benefit
> of this function in your environment ??

So the benefit is not really in the optimization, although that's a side effect.

Let me clarify the point in the changelog.
Imagine we have these cgroups:


                 A  (usage = 2, limit = 2)
                 |
                / \
               /   \
              /     \
             /       \
            /         \
           /           \
          B             C  (usage = 1, limit = 2)
(usage = 1, limit = 2)


The usage in A is the accumulation of the usage in B and C.
Imagine i want to move a task from C to B. This should work well.
We need to first check if we can charge B and do it, and then later
uncharge C.

But if we do:

	err = res_counter_charge(B)
	if (err)
		exit
	res_counter_uncharge(C)

it is going to fail because charging B will also charge A. And A
will refuse because it's already full. Ideally we should first uncharge
C and then charge B, so that A doesn't reject:

	res_counter_uncharge(C)
	err = res_countrer_charge(B)
	if (err)
		res_counter_charge(C)

The problem is that if charging B fails we need to rollback on C, but it might
be too late as a fork might have happen inside C since we uncharged it, so we couldn't
charge it back.

So the only solution is to first charge B but stop the charge propagation on A.
And then uncharge on C but stop uncharge on A.