[Devel] Re: [PATCH] devcgroup: avoid using cgroup_lock

[Li Zefan]

Serge E. Hallyn wrote:
> Quoting Serge E. Hallyn (serue at us.ibm.com):
>> Quoting Li Zefan (lizf at cn.fujitsu.com):
>>>>> @@ -426,11 +431,11 @@ static int devcgroup_access_write(struct cgroup *cgrp, struct cftype *cft,
>>>>>  				  const char *buffer)
>>>>>  {
>>>>>  	int retval;
>>>>> -	if (!cgroup_lock_live_group(cgrp))
>>>> Does it matter that we no longer check for cgroup_is_removed()?
>>>>
>>> No, this means in a rare case that the write handler is called when the cgroup
>>> is dead, we still do the update work instead of returning ENODEV.
>>>
>>> This is ok, since at that time, accessing cgroup and devcgroup is still valid,
>>> but will have no effect since there is no task in this cgroup and the cgroup
>>> will be destroyed soon.
>> Ok, just wanted to make sure the devcgroup couldn't be partially torn
>> down and risking NULL or freed-memory derefs...
> 
> Ok, so the cgroup's files will be deleted first, then on the directory
> removal the cgroup's data (each whitelist entry) is deleted.  So we can
> let that ordering (by cgroup_clear_directory) ensure that nothing inside
> a file write can happen while the destroy handler is called, right?
> 

When we are in the read/write handler, we have a pin in the dir's dentry
(dentry->d_count > 0), thus cgroup_diput() which destroys the cgroup won't
be called during the read/write.

> (That's why I was worried about not using the cgroup_lock: we need some
> way of synchronizing those.  But I guess we're fine)
> 

Many read/write handlers in other cgroup subsystems don't take cgroup_lock.
:)

>> BTW is that against linux-next?  (didn't seem to apply cleanly against
>> my 2.6.29-rc9)  I guess I'd like to do a little test before acking,
>> though it looks ok based on your answer.
> 
> Acked-by: Serge Hallyn <serue at us.ibm.com>
> 
> -serge
> 
> 


_______________________________________________
Containers mailing list
Containers at lists.linux-foundation.org
https://lists.linux-foundation.org/mailman/listinfo/containers