Re: [PATCH 2/3] mm, memcg: inline mem_cgroup_{charge/uncharge} to improve disabled memcg config

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



On Fri, Jul 9, 2021 at 8:18 AM Suren Baghdasaryan <surenb@xxxxxxxxxx> wrote:
>
> On Fri, Jul 9, 2021 at 7:48 AM Johannes Weiner <hannes@xxxxxxxxxxx> wrote:
> >
> > On Thu, Jul 08, 2021 at 05:05:08PM -0700, Suren Baghdasaryan wrote:
> > > Inline mem_cgroup_{charge/uncharge} and mem_cgroup_uncharge_list functions
> > > functions to perform mem_cgroup_disabled static key check inline before
> > > calling the main body of the function. This minimizes the memcg overhead
> > > in the pagefault and exit_mmap paths when memcgs are disabled using
> > > cgroup_disable=memory command-line option.
> > > This change results in ~0.4% overhead reduction when running PFT test
> > > comparing {CONFIG_MEMCG=n} against {CONFIG_MEMCG=y, cgroup_disable=memory}
> > > configurationon on an 8-core ARM64 Android device.
> > >
> > > Signed-off-by: Suren Baghdasaryan <surenb@xxxxxxxxxx>
> >
> > Sounds reasonable to me as well. One comment:
> >
> > > @@ -693,13 +693,59 @@ static inline bool mem_cgroup_below_min(struct mem_cgroup *memcg)
> > >               page_counter_read(&memcg->memory);
> > >  }
> > >
> > > -int mem_cgroup_charge(struct page *page, struct mm_struct *mm, gfp_t gfp_mask);
> > > +struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm);
> > > +
> > > +int __mem_cgroup_charge(struct page *page, struct mem_cgroup *memcg,
> > > +                     gfp_t gfp);
> > > +/**
> > > + * mem_cgroup_charge - charge a newly allocated page to a cgroup
> > > + * @page: page to charge
> > > + * @mm: mm context of the victim
> > > + * @gfp_mask: reclaim mode
> > > + *
> > > + * Try to charge @page to the memcg that @mm belongs to, reclaiming
> > > + * pages according to @gfp_mask if necessary. if @mm is NULL, try to
> > > + * charge to the active memcg.
> > > + *
> > > + * Do not use this for pages allocated for swapin.
> > > + *
> > > + * Returns 0 on success. Otherwise, an error code is returned.
> > > + */
> > > +static inline int mem_cgroup_charge(struct page *page, struct mm_struct *mm,
> > > +                                 gfp_t gfp_mask)
> > > +{
> > > +     struct mem_cgroup *memcg;
> > > +     int ret;
> > > +
> > > +     if (mem_cgroup_disabled())
> > > +             return 0;
> > > +
> > > +     memcg = get_mem_cgroup_from_mm(mm);
> > > +     ret = __mem_cgroup_charge(page, memcg, gfp_mask);
> > > +     css_put(&memcg->css);
> > > +
> > > +     return ret;
> >
> > Why not do
> >
> > int __mem_cgroup_charge(struct page *page, struct mm_struct *mm,
> >                         gfp_t gfp_mask);
> >
> > static inline int mem_cgroup_charge(struct page *page, struct mm_struct *mm,
> >                                     gfp_t gfp_mask)
> > {
> >         if (mem_cgroup_disabled())
> >                 return 0;
> >
> >         return __mem_cgroup_charge(page, memcg, gfp_mask);
> > }
> >
> > like in the other cases as well?
> >
> > That would avoid inlining two separate function calls into all the
> > callsites...
> >
> > There is an (internal) __mem_cgroup_charge() already, but you can
> > rename it charge_memcg().
>
> Sounds good. I'll post an updated version with your suggestion.
> Thanks for the review, Johannes!

Posted v2 just for this patch at
https://lore.kernel.org/patchwork/patch/1455550 . Please let me know
if you want me to resent the whole patchset instead of just this
patch.
>
> >




[Index of Archives]     [Linux ARM Kernel]     [Linux ARM]     [Linux Omap]     [Fedora ARM]     [IETF Annouce]     [Bugtraq]     [Linux OMAP]     [Linux MIPS]     [eCos]     [Asterisk Internet PBX]     [Linux API]

  Powered by Linux