Re: [RFC PATCH] cgroup/cpuset: fix a memory binding failure for cgroup v2

[Waiman Long <longman@redhat.com>]

On 4/24/22 12:04, Waiman Long wrote:
> On 4/21/22 18:22, Tejun Heo wrote:
>> cc'ing Waiman and copying the whole body.
>>
>> Waiman, can you please take a look?
>>
>> Thanks.
>>
>> On Tue, Apr 19, 2022 at 10:09:58AM +0800, Feng Tang wrote:
>>> We got report that setting cpuset.mems failed when the nodemask
>>> contains a newly onlined memory node (not enumerated during boot)
>>> for cgroup v2, while the binding succeeded for cgroup v1.
>>>
>>> The root cause is, for cgroup v2, when a new memory node is onlined,
>>> top_cpuset's 'mem_allowed' is not updated with the new nodemask of
>>> memory nodes, and the following setting memory nodemask will fail,
>>> if the nodemask contains a new node.
>>>
>>> Fix it by updating top_cpuset.mems_allowed right after the
>>> new memory node is onlined, just like v1.
>>>
>>> Signed-off-by: Feng Tang <feng.tang@intel.com>
>>> ---
>>> Very likely I missed some details here, but it looks strange that
>>> the top_cpuset.mem_allowed is not updatd even after we onlined
>>> several memory nodes after boot.
>>>
>>>   kernel/cgroup/cpuset.c | 3 +--
>>>   1 file changed, 1 insertion(+), 2 deletions(-)
>>>
>>> diff --git a/kernel/cgroup/cpuset.c b/kernel/cgroup/cpuset.c
>>> index 9390bfd9f1cd..b97caaf16374 100644
>>> --- a/kernel/cgroup/cpuset.c
>>> +++ b/kernel/cgroup/cpuset.c
>>> @@ -3314,8 +3314,7 @@ static void cpuset_hotplug_workfn(struct 
>>> work_struct *work)
>>>       /* synchronize mems_allowed to N_MEMORY */
>>>       if (mems_updated) {
>>>           spin_lock_irq(&callback_lock);
>>> -        if (!on_dfl)
>>> -            top_cpuset.mems_allowed = new_mems;
>>> +        top_cpuset.mems_allowed = new_mems;
>>>           top_cpuset.effective_mems = new_mems;
>>>           spin_unlock_irq(&callback_lock);
>>>           update_tasks_nodemask(&top_cpuset);
> The on_dfl check was added by commit 7e88291beefb ("cpuset: make 
> cs->{cpus, mems}_allowed as user-configured masks"). This is the 
> expected behavior for cgroup v2 as we don't want to remove a node 
> because it is hot-removed. However, I do see a problem in case we are 
> adding a node that is not originally in top_cpuset.mems_allowed. We 
> should be allowed to add the extra memory node. So something like
>
>         if (!on_dfl)
>                 top_cpuset.mems_allowed = new_mems;
>         else if (!nodes_subset(new_mems, top_cpuset.mems_allowed))
>                 nodes_or(top_cpuset.mems_allowed, 
> top_cpuset.mems_allowed, new_mems);
>
> For v2, top_cpuset.mems_allowed is set to node_possible_map in 
> cpuset_bind(). Perhaps node_possible_map may not include all the nodes 
> that are hot-pluggable.
>
> I don't know if that is similar problem with cpu_possible_mask or not.


Ah, I know why the top_cpuset.mems_allowed isn't correctly set. There 
are 2 places where it is set:

  - cpuset_bind(): top_cpuset.mems_allowed = node_possible_map
  - cpuset_init_smp(): top_cpuset.mems_allowed = node_states[N_MEMORY];

The first one is correct, but the second isn't. It turns out that 
cpuset_init_smp() can be called later than cpuset_bind().

[    2.611207] cpuset_bind called
[    2.631182] cblist_init_generic: Setting adjustable number of 
callback queues.
[    3.082357] cpuset_init_smp called

So the proper fix may be to make sure that top_cpuset.mems_allowed is 
initialized correctly.

Cheers,
Longman