Re: Kernel crash in cgroup_pidlist_destroy_work_fn()

[Li Zefan <lizefan-hv44wF8Li93QT0dZR+AlfA@public.gmane.org>]

On 2014/9/17 13:29, Li Zefan wrote:
> On 2014/9/17 7:56, Cong Wang wrote:
>> Hi, Tejun
>>
>>
>> We saw some kernel null pointer dereference in
>> cgroup_pidlist_destroy_work_fn(), more precisely at
>> __mutex_lock_slowpath(), on 3.14. I can show you the full stack trace
>> on request.
>>
> 
> Yes, please.
> 
>> Looking at the code, it seems flush_workqueue() doesn't care about new
>> incoming works, it only processes currently pending ones, if this is
>> correct, then we could have the following race condition:
>>
>> cgroup_pidlist_destroy_all():
>>         //...
>>         mutex_lock(&cgrp->pidlist_mutex);
>>         list_for_each_entry_safe(l, tmp_l, &cgrp->pidlists, links)
>>                 mod_delayed_work(cgroup_pidlist_destroy_wq,
>> &l->destroy_dwork, 0);
>>         mutex_unlock(&cgrp->pidlist_mutex);
>>
>>         // <--- another process calls cgroup_pidlist_start() here
>> since mutex is released
>>
>>         flush_workqueue(cgroup_pidlist_destroy_wq); // <--- another
>> process adds new pidlist and queue work in pararell
>>         BUG_ON(!list_empty(&cgrp->pidlists)); // <--- This check is
>> passed, list_add() could happen after this
>>
> 
> Did you confirm this is what happened when the bug was triggered?
> 
> I don't think the race condition you described exists. In 3.14 kernel,
> cgroup_diput() won't be called if there is any thread running
> cgroup_pidlist_start(). This is guaranteed by vfs.
> 
> But newer kernels are different. Looks like the bug exists in those
> kernels.
> 

Newer kernels should be also fine.

If cgroup_pidlist_destroy_all() is called, it means kernfs has already
removed the tasks file, and even if you still have it opened, when
you try to read it, it will immediately return an errno.

                                fd = open(cgrp/tasks)
cgroup_rmdir(cgrp)
  cgroup_destroy_locked(c)
    kernfs_remove()
  ...
css_free_work_fn()
  cgroup_pidlist_destroy_all()
                               read(fd of cgrp/tasks)
                                 return -ENODEV

So cgroup_pidlist_destroy_all() won't race with cgroup_pidlist_start().