Re: [PATCH v5 4/5] cpusets: Update tasks' cpus_allowed mask upon updates to root cpuset

["Srivatsa S. Bhat" <srivatsa.bhat@linux.vnet.ibm.com>]

On 05/24/2012 02:43 PM, Peter Zijlstra wrote:

> On Thu, 2012-05-17 at 22:33 +0530, Srivatsa S. Bhat wrote:
>> +               /*
>> +                * Restore only the top_cpuset because it has to track
>> +                * cpu_active_mask always.
>> +                * (We don't need to do anything if we come here during resume
>> +                * from suspend, since top_cpuset.cpus_allowed will already be
>> +                * equal to cpu_active_mask.)
>> +                */
>> +               if (root == &top_cpuset && !cpumask_equal(root->cpus_allowed,
>> +                                                         cpu_active_mask)) {
>> +                       mutex_lock(&callback_mutex);
>> +                       cpumask_copy(root->cpus_allowed, cpu_active_mask);
>> +                       mutex_unlock(&callback_mutex);
>> +                       update_tasks_cpumask(root, NULL);
>> +               } 
> 
> This looks absolutely broken.
> 
> Suppose I set an explicit cpu affinity mask on my task, which per not
> using cpusets is in the root group.
> 
> If I then do a hotplug cycle, I'll find my affinity mask is lost.
> 


Sorry, my bad, I hadn't considered that. Thanks for pointing it out!

So, I am wondering how we ought to deal with CPU hotplug for tasks attached
to the root cpuset..

Considering tasks attached to the root cpuset, if a cpu present in a task's
cpus_allowed mask goes offline, it should be removed from that mask right?
And if that cpu comes back online, it should not be put back to the task's
cpus_allowed mask (just like we don't put back cpus in non-root cpusets).

Is the above understanding correct?

In the current kernel, during cpu hotplug, we don't touch cpus_allowed mask
of the tasks attached to the root cpuset at all.. Whereas we update the
cpus_allowed mask of tasks belonging to non-root cpusets, during cpu offline.

So, is this differentiation intended?

Regards,
Srivatsa S. Bhat