Re: [-mm] Add an owner to the mm_struct (v7)

[Balbir Singh <balbir@linux.vnet.ibm.com>]

Paul Menage wrote:
> On Thu, Apr 3, 2008 at 11:11 AM, Balbir Singh <balbir@linux.vnet.ibm.com> wrote:
>>  >
>>  > Hmm, is this new check for delay_group_leader() safe? Won't we have
>>  > called exit_cgroup() by this point, and hence be reassigned to the
>>  > root cgroup? And so mm->owner->cgroups won't point to the right place?
>>  >
>>
>>  cgroup_exit() comes in much later after exit_mm(). Moreover delay_group_leader()
>>  is a function that checks to see if
> 
> Sorry, I was unclear.
> 
> Yes, the call to cgroup_exit() comes much later than exit_mm() - but
> it probably does come before the other users of the mm have finished
> using the mm. So can't we end up with a situation like this?
> 
> A (group leader) exits; at this point, A->mm->owner == A
> A calls exit_mm(), sees delay_group_leader(), doesn't change A->mm->owner
> A calls cgroup_exit(), A->cgroups is set to init_css_set.
> B (another thread) does something with B->mm->owner->cgroups (e.g. VM
> accounting) and accesses the wrong group

Hi, Paul,

That is indeed quite bad. Do we have to retire the group_leader to init_css_set?
Can we not check for delay_group_leader() there?

-- 
	Warm Regards,
	Balbir Singh
	Linux Technology Center
	IBM, ISTL