Re: BUG: HANG_DETECT waiting for migration_cpu_stop() complete

[Jing-Ting Wu]

Hi, Mukesh



https://lore.kernel.org/lkml/YvrWaml3F+x9Dk+T@xxxxxxxxxxxxxxx/ is for
fix cgroup_threadgroup_rwsem <-> cpus_read_lock() deadlock.
But this issue is cgroup_threadgroup_rwsem <-> cpuset_rwsem deadlock.

I think they are not same issue.
Do the patch is useful for this issue?



Best regards,
Jing-Ting Wu


On Mon, 2022-09-05 at 12:14 +0530, Mukesh Ojha wrote:
> This is fixed by this.
> 
> https://lore.kernel.org/lkml/YvrWaml3F+x9Dk+T@xxxxxxxxxxxxxxx/
> 
> -Mukesh
> 
> On 9/5/2022 8:17 AM, Jing-Ting Wu wrote:
> > Hi,
> > 
> > We meet the HANG_DETECT happened in T SW version with kernel-5.15.
> > Many tasks have been blocked for a long time.
> > 
> > 
> > Root cause:
> > migration_cpu_stop() is not complete due to
> > is_migration_disabled(p) is
> > true, complete is false and complete_all() never get executed.
> > It let other task wait the rwsem.
> > 
> > Detail:
> > system_server waiting for cgroup_threadgroup_rwsem.
> > OomAdjuster is holding the cgroup_threadgroup_rwsem and waiting for
> > cpuset_rwsem.
> > cpuset_hotplug_workfn is holding the cpuset_rwsem and waiting for
> > affine_move_task() complete.
> > affine_move_task() waiting for migration_cpu_stop() complete.
> > 
> > The backtrace of system_server:
> > __switch_to
> > __schedule
> > schedule
> > percpu_rwsem_wait
> > __percpu_down_read
> > cgroup_css_set_fork => wait for cgroup_threadgroup_rwsem
> > cgroup_can_fork
> > copy_process
> > kernel_clone
> > 
> > The backtrace of OomAdjuster:
> > __switch_to
> > __schedule
> > schedule
> > percpu_rwsem_wait
> > percpu_down_write
> > cpuset_can_attach => wait for cpuset_rwsem
> > cgroup_migrate_execute
> > cgroup_attach_task
> > __cgroup1_procs_write => hold cgroup_threadgroup_rwsem
> > cgroup1_procs_write
> > cgroup_file_write
> > kernfs_fop_write_iter
> > vfs_write
> > ksys_write
> > 
> > The backtrace of cpuset_hotplug_workfn:
> > __switch_to
> > __schedule
> > schedule
> > schedule_timeout
> > wait_for_common
> > affine_move_task => wait for complete
> > __set_cpus_allowed_ptr_locked
> > update_tasks_cpumask
> > cpuset_hotplug_update_tasks => hold cpuset_rwsem
> > cpuset_hotplug_workfn
> > process_one_work
> > worker_thread
> > kthread
> > 
> > 
> > In affine_move_task() will call migration_cpu_stop() and wait for
> > it
> > complete.
> > In normal case, if migration_cpu_stop() complete it will inform
> > everyone that he is done.
> > But there is an exception case that will not notify.
> > If is_migration_disabled(p) is true and complete will always is
> > false,
> > then complete_all() never get executed.
> > 
> > static int migration_cpu_stop(void *data)
> > {
> > ...
> >      bool complete = false;
> > ...
> > 
> >      if (task_rq(p) == rq) {
> >          if (is_migration_disabled(p))
> >                goto out; => is_migration_disabled(p) = true,
> >                             so complete = false.
> >              ...
> >          }
> > ...
> > 
> > out:
> > ...
> >      if (complete) => complete = false,
> >                       so complete_all() never get executed.
> >          complete_all(&pending->done);
> > 
> >          return 0;
> > }
> > 
> > 
> > Review the code, we found that there are many places can change
> > is_migration_disabled() value.
> > (such as: __rt_spin_lock(), rt_read_lock(), rt_write_lock(), ...)
> > 
> > Do you have any suggestion for this issue?
> > Thank you.
> > 
> > 
> > 
> > 
> > Best regards,
> > Jing-Ting Wu
> > 
> >