Re: [patch 0/3] mm: memcontrol: eliminate charge reparenting

[Vladimir Davydov <vdavydov@parallels.com>]

Hi Johannes,

On Sat, Sep 20, 2014 at 04:00:32PM -0400, Johannes Weiner wrote:
> The decoupling of css from the user-visible cgroup, word-sized per-cpu
> css reference counters, and css iterators that include offlined groups
> means we can take per-charge css references, continue to reclaim from
> offlined groups, and so get rid of the error-prone charge reparenting.

I haven't reviewed this set yet, but I agree that zapping user memory
reparenting sounds like a sane idea, because reparenting won't let the
css go in most cases anyway due to swap and kmem charges.

However, I think we must reparent list_lru items, otherwise per-memcg
arrays (kmem_caches, list_lrus) will grow uncontrollably due to dead
css's, which is unacceptable. Note it isn't the same as the user memory
reparenting, because we don't need to reparent kmem_cache objects or
charges - they can stay where they are pinning the css till they are
freed, because the memcg_cache_id, which I want to free on offline, is
not used for kmem allocations/frees after css offline. Actually we only
need to empty the list_lru corresponding to the dead memory cgroup,
which is relatively easy to implement. This is what patch 13 of the "Per
memcg slab shrinkers" patch set, which I sent recently, does (see
https://lkml.org/lkml/2014/9/21/64).

What do you think about it?

Thanks,
Vladimir

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>