BUG: HANG_DETECT waiting for migration_cpu_stop() complete

[Waiman Long]

On 9/6/22 16:01, Waiman Long wrote:
> On 9/6/22 14:30, Tejun Heo wrote:
>> Hello,
>>
>> (cc'ing Waiman in case he has a better idea)
>>
>> On Mon, Sep 05, 2022 at 04:22:29PM +0800, Jing-Ting Wu wrote:
>>> https://lore.kernel.org/lkml/YvrWaml3F+x9Dk+T@slm.duckdns.org/ is for
>>> fix cgroup_threadgroup_rwsem <-> cpus_read_lock() deadlock.
>>> But this issue is cgroup_threadgroup_rwsem <-> cpuset_rwsem deadlock.
>> If I'm understanding what you're writing correctly, this isn't a 
>> deadlock.
>> The cpuset_hotplug_workfn simply isn't being woken up while holding
>> cpuset_rwsem and others are just waiting for that lock to be released.
>
> I believe it is probably a bug in the scheduler core code. 
> __set_cpus_allowed_ptr_locked() calls affine_move_task() to move to a 
> random cpu within the new set allowable CPUs. However, if migration is 
> disabled, it shouldn't call affine_move_task() at all. Instead, I 
> would suggest that if the current cpu is within the new allowable 
> cpus, it should just skip doing affine_move_task(). Otherwise, it 
> should fail __set_cpus_allowed_ptr_locked().

Maybe like

diff --git a/kernel/sched/core.c b/kernel/sched/core.c
index 838623b68031..5d9ea1553ec0 100644
--- a/kernel/sched/core.c
+++ b/kernel/sched/core.c
@@ -2794,9 +2794,9 @@ static int __set_cpus_allowed_ptr_locked(struct 
task_struct *p,
                 if (cpumask_equal(&p->cpus_mask, new_mask))
                         goto out;

-               if (WARN_ON_ONCE(p == current &&
-                                is_migration_disabled(p) &&
-                                !cpumask_test_cpu(task_cpu(p), 
new_mask))) {
+               if (is_migration_disabled(p) &&
+                   !cpumask_test_cpu(task_cpu(p), new_mask)) {
+                       WARN_ON_ONCE(p == current);
                         ret = -EBUSY;
                         goto out;
                 }
@@ -2818,7 +2818,11 @@ static int __set_cpus_allowed_ptr_locked(struct 
task_struct *p,
         if (flags & SCA_USER)
                 user_mask = clear_user_cpus_ptr(p);

-       ret = affine_move_task(rq, p, rf, dest_cpu, flags);
+       if (!is_migration_disabled(p) || (flags & SCA_MIGRATE_ENABLE)) {
+               ret = affine_move_task(rq, p, rf, dest_cpu, flags);
+       } else {
+               task_rq_unlock(rq, p, rf);
+       }

         kfree(user_mask);

I haven't tested it myself, though.

Cheers,
Longman