On 10/1/21 4:24 PM, Shakeel Butt wrote:
On Fri, Oct 1, 2021 at 12:10 PM Waiman Long <longman@xxxxxxxxxx> wrote:
To ensure that a to-be-offlined memcg fetched from objcg remains
valid (has non-zero reference count) within a RCU critical section,
a synchronize_rcu() call is inserted at the end of memcg_offline_kmem().
With that change, we no longer need to use css_tryget()
in get_mem_cgroup_from_objcg() as the final css_put() in
css_killed_work_fn() would not have been called yet.
The obj_cgroup_uncharge_pages() function is simplifed to perform
the whole uncharge operation within a RCU critical section saving a
css_get()/css_put() pair.
Signed-off-by: Waiman Long <longman@xxxxxxxxxx>
---
mm/memcontrol.c | 16 +++++++++++-----
1 file changed, 11 insertions(+), 5 deletions(-)
diff --git a/mm/memcontrol.c b/mm/memcontrol.c
index 8177f253a127..1dbb37d96e49 100644
--- a/mm/memcontrol.c
+++ b/mm/memcontrol.c
@@ -2769,10 +2769,8 @@ static struct mem_cgroup *get_mem_cgroup_from_objcg(struct obj_cgroup *objcg)
struct mem_cgroup *memcg;
rcu_read_lock();
-retry:
memcg = obj_cgroup_memcg(objcg);
- if (unlikely(!css_tryget(&memcg->css)))
- goto retry;
+ css_get(&memcg->css);
rcu_read_unlock();
return memcg;
@@ -2947,13 +2945,14 @@ static void obj_cgroup_uncharge_pages(struct obj_cgroup *objcg,
{
struct mem_cgroup *memcg;
- memcg = get_mem_cgroup_from_objcg(objcg);
+ rcu_read_lock();
+ memcg = obj_cgroup_memcg(objcg);
if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
page_counter_uncharge(&memcg->kmem, nr_pages);
refill_stock(memcg, nr_pages);
- css_put(&memcg->css);
+ rcu_read_unlock();
}
/*
@@ -3672,6 +3671,13 @@ static void memcg_offline_kmem(struct mem_cgroup *memcg)
memcg_drain_all_list_lrus(kmemcg_id, parent);
memcg_free_cache_id(kmemcg_id);
+
+ /*
+ * To ensure that a to-be-offlined memcg fetched from objcg remains
+ * valid within a RCU critical section, we need to wait here until
+ * the a grace period has elapsed.
+ */
+ synchronize_rcu();
This is called with cgroup_mutex held from css_offline path and
synchronize_rcu() can be very expensive on a busy system, so, this
will indirectly impact all the code paths which take cgroup_mutex.
Yes, you are right. Just don't consider this patch for the time being. I
will need to find a way to work around that.
Thanks,
Longman
