[PATCH 2/8] cgroup: kill CSS_REMOVED

Glauber Costa glommer at parallels.com
Wed Oct 31 17:19:51 UTC 2012


On 10/31/2012 09:10 PM, Tejun Heo wrote:
> Hello, Glauber.
> 
> On Wed, Oct 31, 2012 at 10:06 AM, Glauber Costa <glommer at parallels.com> wrote:
>> This is not the topic of this thread, but since you brought it:
>> If you take a look at the description patch in the patch I sent, the
>> problem I outlined is that at create time, we don't know anything about
>> which will the css_id be - and I would like to avoid creating yet
>> another index.
>>
>> Is there any way you would suggest of handling this ? Any chance of us
>> allocating the css_id earlier then?
> 
> I only glanced the patch description but the problem I'm trying to
> solve is locking - currently we don't have a place where a controller
> can tell a cgroup is becoming online; thus it has nothing to
> synchronized against and tell that a cgroup is alive or not.  As for
> css_id allocation, maybe you can deal with that in ->post_create() or
> maybe we can allocate css_id earlier (but where would it be stored?).
> I'll look into it.
> 

The css_id is allocated right after ->create(). So if post_create() is
called after that, I can use it just fine - which is what I do in my
patch. Problem is, since I do memory allocation based on that, it can fail.

So although I would have to look at your series myself to see exactly
what you are trying to achieve (looking forward), I seemed natural to me
to think about it terms of "early_create" + "late_create" (->create())
and ->post_create()), where the later is a callback when things are
readier (in this case, the css_id).

I don't see post_create failing as a huge problem. The natural
synchronization point would be "right after post_create" - then you can
definitely tell that it is online. Although this can be viewed a bit as
"exposing internals", creating is different then destroying: When you
create, you may not have all data yet. When destroying, you do - and
want to get rid of it. So this kind of bootstrapping is pretty standard
and common.





More information about the Containers mailing list