Re: [PATCH] memcg: remove unused mem_cgroup_from_obj()

[Yosry Ahmed]

On Sun, May 28, 2023 at 6:02 AM Muchun Song <muchun.song@xxxxxxxxx> wrote:
>
>
>
> > On May 28, 2023, at 02:54, Yosry Ahmed <yosryahmed@xxxxxxxxxx> wrote:
> >
> > On Sat, May 27, 2023 at 8:07 AM Matthew Wilcox <willy@xxxxxxxxxxxxx> wrote:
> >>
> >> On Fri, May 26, 2023 at 09:13:05PM -0700, Yosry Ahmed wrote:
> >>> On Fri, May 26, 2023 at 9:01 PM Matthew Wilcox <willy@xxxxxxxxxxxxx> wrote:
> >>>>
> >>>> On Sat, May 27, 2023 at 06:31:26PM +0800, Miaohe Lin wrote:
> >>>>> The function mem_cgroup_from_obj() is not used anymore. Remove it and
> >>>>> clean up relevant comments.
> >>>>
> >>>> You should have looked at the git history to see why it was created
> >>>> and who used it.
> >>>>
> >>>> Shakeel, Vasily, are you going to retry adding commit 1d0403d20f6c?
> >>>
> >>> That commit did not introduce the function though, no? It was
> >>> introduced before it and replaced by other variants over time (like
> >>> mem_cgroup_from_slab_obj()). It looks like that commit was reverted ~9
> >>> months ago. We can always bring it back if/when needed.
> >>
> >> The commit immediately preceding it is fc4db90fe71e.
> >>
> >> Of course we can bring it back.  It's just code.  But avoiding
> >> unnecessary churn is also good.  Let's wait to hear from Vasily.
> >>
> >>> It also looks to me that 1d0403d20f6c was using mem_cgroup_from_obj()
> >>> on a struct net object, which is allocated in net_alloc() from a slab
> >>> cache, so mem_cgroup_from_slab_obj() should be sufficient, no?
> >>
> >> Clearly not.
> >
> > I dived deeper into the history on LKML, and you are right:
> > https://lore.kernel.org/all/Yp4F6n2Ie32re7Ed@qian/
> >
> > I still do not understand why mem_cgroup_from_slab_obj() would not be
> > sufficient, so I am hoping Vasily or Shakeel can help me understand
> > here. Seems to be something arch-specific.
>
> I think it is because *init_net* which is not allocated from slab meant
> its address does not belong to linear mapping addresses on arm64. However,
> virt_to_page() is only applicable to linear mapping addresses. So,
> mem_cgroup_from_slab_obj() is not sufficient. mem_cgroup_from_obj() is used
> in this case, which will use vmalloc_to_page() for the page associated
> with *init_net*. If Vasily does not want to bring commit 1d0403d20f6c back,
> this patch LGTM. Otherwise, let's wait for Vasily.

I see, thanks for the context, Muchun!

>
> Thanks.
>