Re: [PATCH 1/3] vmpressure: document why css_get/put is not necessary for work queue based signaling

On Fri 12-07-13 11:48:36, Tejun Heo wrote:

On Fri, Jul 12, 2013 at 11:24:56AM +0200, Michal Hocko wrote:

quoted

Cgroup events are unregistered from the workqueue context by
cgroup_event_remove scheduled by cgroup_destroy_locked (when a cgroup is
removed by rmdir).

cgroup_event_remove removes the eventfd wait queue from the work
queue, then it unregisters all the registered events and finally
puts a reference to the cgroup dentry. css_free which triggers memcg
deallocation is called after the last reference is dropped.

The scheduled vmpressure work item either happens before
cgroup_event_remove or it is not triggered at all so it always happen
_before_ the last dput thus css_free.

I don't follow what the above has to do with ensuring work item
execution is finished before the underlying data structure is
released.  How are the above relevant?  What am I missing here?

OK, it seems I managed to confuse myself. I thought that 
remove_wait_queue(event->wqh, &event->wait) called from
cgroup_event_remove guarantee that vmpr event would go away with that
workqueue. But now that I am looking at it, vmpr->work seems to be
living in a completely independent queue.

quoted

This patch just documents this trickiness.

This doesn't have to be tricky at all.  It's a *completely* routine
thing.  Would you please stop making it one?

Fair enough. I will repost the series shortly.
-- 
Michal Hocko
SUSE Labs