Re: [PATCH] cgroups: fix probable race with put_css_set[_taskexit] and find_css_set

[Li Zefan <lizf@cn.fujitsu.com>]

Greg KH wrote:
> On Tue, Sep 09, 2008 at 10:31:24PM -0700, Paul Menage wrote:
>> On Tue, Sep 9, 2008 at 10:01 PM, Greg KH <greg@kroah.com> wrote:
>>> What are you trying to solve here with this change?  I agree, it does
>>> seem a bit "chaotic" :)
>> There's a place in cgroups that uses kref_put() to release an object;
>> the release function *then* takes a write-lock and removes the object
>> from a lookup table; it could race with another thread that searches
>> the lookup table (while holding a read-lock) and does kref_get() on
>> the same object.
> 
> Ick, yeah that's not good.
> 
> What about the way everyone else solves this, grab the lock before you
> call kref_put()?
> 

do_exit()
  cgroup_exit()
    put_css_set_taskexit()
      kref_put()

If we grab the lock before kref_put(), we add overhead to do_exit(), which
is what we are trying to avoid here.

>> The current fix is for the release function to recheck inside the lock
>> that the object's refcount is still zero, and only actually
>> unlink/free it if so. And actually I've just realised that this isn't
>> actually even safe, since the thread that just acquired the object
>> could kref_put() it almost immediately, which would leave two threads
>> both trying to unlink/free the object.
> 
> Yeah, don't do that :)
>