Re: [External] Re: [PATCH] mm: memcontrol: remove rcu_read_lock from get_mem_cgroup_from_page

[Muchun Song]

On Fri, Feb 5, 2021 at 6:32 PM Michal Hocko <mhocko@xxxxxxxx> wrote:
>
> On Fri 05-02-21 17:14:30, Muchun Song wrote:
> > On Fri, Feb 5, 2021 at 4:36 PM Michal Hocko <mhocko@xxxxxxxx> wrote:
> > >
> > > On Fri 05-02-21 14:27:19, Muchun Song wrote:
> > > > The get_mem_cgroup_from_page() is called under page lock, so the page
> > > > memcg cannot be changed under us.
> > >
> > > Where is the page lock enforced?
> >
> > Because it is called from alloc_page_buffers(). This path is under
> > page lock.
>
> I do not see any page lock enforecement there. There is not even a
> comment requiring that. Can we grow more users where this is not the
> case? There is no actual relation between alloc_page_buffers and
> get_mem_cgroup_from_page except that the former is the only _current_
> existing user. I would be careful to dictate locking based solely on
> that.

Yeah, there is no comment requiring that. I have seen all the callers
of the alloc_page_buffers. I found that it is under page lock.
But it seems it is not the key point here. I should delete those
comments from the commit log.

>
> > > > Also, css_get is enough because page
> > > > has a reference to the memcg.
> > >
> > > tryget used to be there to guard against offlined memcg but we have
> > > concluded this is impossible in this path. tryget stayed there to catch
> > > some unexpected cases IIRC.
> >
> > Yeah, it can catch some unexpected cases. But why is this path
> > special so that we need a tryget?
>
> I do not remember details and the changelog of that change is not
> explicit but I suspect it was just because this one could trigger as
> there are external callers to memcg. Is this protection needed? I am not
> sure, this is for you to justify if you want to remove it.

I am sure it is not needed.

>
> > > > If we really want to make the get_mem_cgroup_from_page() suitable for
> > > > arbitrary page, we should use page_memcg_rcu() instead of page_memcg()
> > > > and call it after rcu_read_lock().
> > >
> > > What is the primary motivation to change this code? is the overhead of
> > > tryget/RCU something that needs optimizing?
> >
> > Actually, the rcu_read_lock() is not necessary here. So it is better to
> > remove it (indeed reduce some code).
>
> Please state your reasoning in the changelog including benefits we get
> from it.

OK.

> --
> Michal Hocko
> SUSE Labs