[Devel] Re: [ckrm-tech] [PATCH 1/7] containers (V7): Generic container system abstracted from cpusets code

[Srivatsa Vaddagiri]

On Sat, Mar 24, 2007 at 06:41:28PM -0700, Paul Jackson wrote:
> > the following code becomes racy with cpuset_exit() ...
> > 
> >         atomic_inc(&cs->count);
> >         rcu_assign_pointer(tsk->cpuset, cs);
> >         task_unlock(tsk);
> 
> eh ... so ... ?
> 
> I don't know of any sequence where that causes any problem.
> 
> Do you see one?

Let's say we had two cpusets CS1 amd CS2 (both different from top_cpuset).
CS1 has just one task T1 in it (CS1->count = 0) while CS2 has no tasks
in it (CS2->count = 0).

Now consider:

--------------------------------------------------------------------
CPU0 (attach_task T1 to CS2)			CPU1 (T1 is exiting)
--------------------------------------------------------------------

task_lock(T1);

oldcs = tsk->cpuset;
[oldcs = CS1]

T1->flags & PF_EXITING? (No)			

						T1->flags = PF_EXITING;

atomic_inc(&CS2->count);

						cpuset_exit()
						    cs = tsk->cpuset; (cs = CS1)
							
T1->cpuset = CS2;

						    T1->cpuset = &top_cpuset;

task_unlock(T1);


CS2 has one bogus count now (with no tasks in it), which may prevent it from 
being removed/freed forever.


Not just this, continuing further we have more trouble:

--------------------------------------------------------------------
CPU0 (attach_task T1 to CS2)			CPU1 (T1 is exiting)
--------------------------------------------------------------------

synchronize_rcu()
						    atomic_dec(&CS1->count);
						    [CS1->count = 0]

if atomic_dec_and_test(&oldcs->count))
	[CS1->count = -1]



We now have CS1->count negative. Is that good? I am uncomfortable ..

We need a task_lock() in cpuset_exit to avoid this race.

-- 
Regards,
vatsa