On Fri 07-02-14 12:04:20, Johannes Weiner wrote:
> Reparenting memory charges in the css_free() callback was meant as a
> temporary fix for charges that race with offlining, but after some
> follow-up discussion, it turns out that this is really the right place
> to reparent charges because it guarantees none are in-flight.
>
> Make clear that the reparenting in css_offline() is an optimistic
> sweep of established charges because swapout records might hold up
> css_free() indefinitely, but that in fact the css_free() reparenting
> is the properly synchronized one.
>
> Signed-off-by: Johannes Weiner <han...@cmpxchg.org>
OK, I am still thinking about 2 stage reparenting. LRU drain part called
from css_offline and charge drain from css_free. But this is a
sufficient for now.
Acked-by: Michal Hocko <mho...@suse.cz>
> ---
> mm/memcontrol.c | 52 +++++++++++++++-------------------------------------
> 1 file changed, 15 insertions(+), 37 deletions(-)
>
> diff --git a/mm/memcontrol.c b/mm/memcontrol.c
> index 639cf58b2643..b8a96c7d1167 100644
> --- a/mm/memcontrol.c
> +++ b/mm/memcontrol.c
> @@ -6600,51 +6600,29 @@ static void mem_cgroup_css_offline(struct
> cgroup_subsys_state *css)
> kmem_cgroup_css_offline(memcg);
>
> mem_cgroup_invalidate_reclaim_iterators(memcg);
> - mem_cgroup_reparent_charges(memcg);
> mem_cgroup_destroy_all_caches(memcg);
> vmpressure_cleanup(&memcg->vmpressure);
> + /*
> + * Memcg gets css references while charging the res_counter,
> + * so we reparent charges in .css_free() when the references
> + * are gone and we know there are no in-flight charges.
> + *
> + * However, at this time, swapout records also hold css refs
> + * indefinitely beyond offlining, which prevent .css_free()
> + * from being called. But after offlining, css_tryget() is
> + * disabled, which means that all the left-over page cache in
> + * the group would be stuck without being reclaimable. Clear
> + * out all those already established charges optimistically
> + * here, and catch any raced charges in .css_free() later on.
> + */
> + mem_cgroup_reparent_charges(memcg);
> }
>
> static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
> {
> struct mem_cgroup *memcg = mem_cgroup_from_css(css);
> - /*
> - * XXX: css_offline() would be where we should reparent all
> - * memory to prepare the cgroup for destruction. However,
> - * memcg does not do css_tryget() and res_counter charging
> - * under the same RCU lock region, which means that charging
> - * could race with offlining. Offlining only happens to
> - * cgroups with no tasks in them but charges can show up
> - * without any tasks from the swapin path when the target
> - * memcg is looked up from the swapout record and not from the
> - * current task as it usually is. A race like this can leak
> - * charges and put pages with stale cgroup pointers into
> - * circulation:
> - *
> - * #0 #1
> - * lookup_swap_cgroup_id()
> - * rcu_read_lock()
> - * mem_cgroup_lookup()
> - * css_tryget()
> - * rcu_read_unlock()
> - * disable css_tryget()
> - * call_rcu()
> - * offline_css()
> - * reparent_charges()
> - * res_counter_charge()
> - * css_put()
> - * css_free()
> - * pc->mem_cgroup = dead memcg
> - * add page to lru
> - *
> - * The bulk of the charges are still moved in offline_css() to
> - * avoid pinning a lot of pages in case a long-term reference
> - * like a swapout record is deferring the css_free() to long
> - * after offlining. But this makes sure we catch any charges
> - * made after offlining:
> - */
> - mem_cgroup_reparent_charges(memcg);
>
> + mem_cgroup_reparent_charges(memcg);
> memcg_destroy_kmem(memcg);
> __mem_cgroup_free(memcg);
> }
> --
> 1.8.5.3
>
--
Michal Hocko
SUSE Labs
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majord...@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
