[Devel] Re: [PATCH v3 16/28] memcg: kmem controller charge/uncharge infrastructure

[Frederic Weisbecker]

On Wed, May 30, 2012 at 04:38:39PM +0400, Glauber Costa wrote:
> On 05/30/2012 04:34 PM, Frederic Weisbecker wrote:
> >On Fri, May 25, 2012 at 05:03:36PM +0400, Glauber Costa wrote:
> >>+bool __mem_cgroup_new_kmem_page(struct page *page, gfp_t gfp)
> >>+{
> >>+	struct mem_cgroup *memcg;
> >>+	struct page_cgroup *pc;
> >>+	bool ret = true;
> >>+	size_t size;
> >>+	struct task_struct *p;
> >>+
> >>+	if (!current->mm || in_interrupt())
> >>+		return true;
> >>+
> >>+	rcu_read_lock();
> >>+	p = rcu_dereference(current->mm->owner);
> >>+	memcg = mem_cgroup_from_task(p);
> >
> >So this takes the memcg of the group owner rather than the
> >task? I understand why we want this for user memory, but for
> >kernel?
> 
> That was already discussed when this first came up in my last submission
> If I recall correctly, Kame pointed out that this would be needed
> for proper OOM-scoring and killing.

Can we have at least a comment in the code that explain the reasons of taking the
owner rather than the task? It's not going to be very obvious to future reviewers.

> Now of course we won't oom kernel threads or anything like that.

Seems we are not even accounting them anyway.

> But since this is also accounted towards memcg, it should at least be
> consistent with each memcg it accounts to.
> 
> We can't account kmem for the thread's memcg, and mem to the process'.

Don't know. This goes a bit against cgroups semantics which group at the task
level and not process. But I personally don't mind much, as long as it's
documented.