Messages in this thread | | | Date | Fri, 16 May 2014 09:28:24 +0800 | From | Li Zefan <> | Subject | Re: [PATCHSET cgroup/for-3.16] cgroup: iterate cgroup_subsys_states directly |
| |
On2014/5/14 21:07, Tejun Heo wrote: > Hello, Li. > > On Wed, May 14, 2014 at 12:21:25PM +0800, Li Zefan wrote: >>> There are now use cases where controllers need to iterate through >>> csses regardless of their online state as long as they have positive >> >> What use cases are we talking about here? > > memcg wants to be able to iterate all csses whose refcnts haven't > reached zero yet so that it can treat offline csses the same way as > online ones in terms of memory reclaim. They don't contain new tasks > so new charges won't be created but offlining won't try to transfer > all charges to the parent but just leave the offline child attached > until all charges are eventually reclaimed from the pressure from the > parent. > > I'm not too familiar with the details but this makes sense in generic > sense too. Offline marks an object starting its draining phase and > release marks the actual destruction point. For controllers with > persistent states like memcg, it's a lot more natural to deal offlined > csses as "active but draining following the usual hierarchical > operation" rather than trying to explicitly update the states from > offline to move them to the parent especially as the effort there > essentially is a waste as most of those moved charges aren't gonna be > used in the parent and will be released eventually. > > Guaranteeing iteration of offline but not-released csses allow > controllers to treat the draining stage between offline and release > more or less identically to online state which in turn can make > ->css_offline() significantly simpler and lighter. >
yeah, fair enough.
| |