Received-SPF: pass (google.com: best guess record for domain of linux-kernel-owner@vger.kernel.org designates 209.132.180.67 as permitted sender) client-ip=209.132.180.67;
MIME-Version: 1.0
References: <20200304022058.248270-1-shakeelb@google.com> <20200305204109.be23f5053e2368d3b8ccaa06@linux-foundation.org>
In-Reply-To: <20200305204109.be23f5053e2368d3b8ccaa06@linux-foundation.org>
From:   Shakeel Butt <shakeelb@google.com>
Date:   Thu, 5 Mar 2020 20:54:39 -0800
Message-ID: <CALvZod7W-Qwa4BRKW0_Ts5f68fwkcqD72SF_4NqZRgEMgA_1-g@mail.gmail.com>
Subject: Re: [PATCH] memcg: optimize memory.numa_stat like memory.stat
To:     Andrew Morton <akpm@linux-foundation.org>
Cc:     Johannes Weiner <hannes@cmpxchg.org>, Roman Gushchin <guro@fb.com>,
        Michal Hocko <mhocko@kernel.org>,
        Linux MM <linux-mm@kvack.org>,
        Cgroups <cgroups@vger.kernel.org>,
        LKML <linux-kernel@vger.kernel.org>
Content-Type: text/plain; charset="UTF-8"
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk

On Thu, Mar 5, 2020 at 8:41 PM Andrew Morton <akpm@linux-foundation.org> wrote:
>
> On Tue,  3 Mar 2020 18:20:58 -0800 Shakeel Butt <shakeelb@google.com> wrote:
>
> > Currently reading memory.numa_stat traverses the underlying memcg tree
> > multiple times to accumulate the stats to present the hierarchical view
> > of the memcg tree. However the kernel already maintains the hierarchical
> > view of the stats and use it in memory.stat. Just use the same mechanism
> > in memory.numa_stat as well.
> >
> > I ran a simple benchmark which reads root_mem_cgroup's memory.numa_stat
> > file in the presense of 10000 memcgs. The results are:
> >
> > Without the patch:
> > $ time cat /dev/cgroup/memory/memory.numa_stat > /dev/null
> >
> > real    0m0.700s
> > user    0m0.001s
> > sys     0m0.697s
> >
> > With the patch:
> > $ time cat /dev/cgroup/memory/memory.numa_stat > /dev/null
> >
> > real    0m0.001s
> > user    0m0.001s
> > sys     0m0.000s
> >
>
> Can't you do better than that ;)
>
> >
> > +     page_state = tree ? lruvec_page_state : lruvec_page_state_local;
> > ...
> >
> > +     page_state = tree ? memcg_page_state : memcg_page_state_local;
> >
>
> All four of these functions are inlined.  Taking their address in this
> fashion will force the compiler to generate out-of-line copies.
>
> If we do it the uglier-and-maybe-a-bit-slower way:
>
> --- a/mm/memcontrol.c~memcg-optimize-memorynuma_stat-like-memorystat-fix
> +++ a/mm/memcontrol.c
> @@ -3658,17 +3658,16 @@ static unsigned long mem_cgroup_node_nr_
>         struct lruvec *lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(nid));
>         unsigned long nr = 0;
>         enum lru_list lru;
> -       unsigned long (*page_state)(struct lruvec *lruvec,
> -                                   enum node_stat_item idx);
>
>         VM_BUG_ON((unsigned)nid >= nr_node_ids);
>
> -       page_state = tree ? lruvec_page_state : lruvec_page_state_local;
> -
>         for_each_lru(lru) {
>                 if (!(BIT(lru) & lru_mask))
>                         continue;
> -               nr += page_state(lruvec, NR_LRU_BASE + lru);
> +               if (tree)
> +                       nr += lruvec_page_state(lruvec, NR_LRU_BASE + lru);
> +               else
> +                       nr += lruvec_page_state_local(lruvec, NR_LRU_BASE + lru);
>         }
>         return nr;
>  }
> @@ -3679,14 +3678,14 @@ static unsigned long mem_cgroup_nr_lru_p
>  {
>         unsigned long nr = 0;
>         enum lru_list lru;
> -       unsigned long (*page_state)(struct mem_cgroup *memcg, int idx);
> -
> -       page_state = tree ? memcg_page_state : memcg_page_state_local;
>
>         for_each_lru(lru) {
>                 if (!(BIT(lru) & lru_mask))
>                         continue;
> -               nr += page_state(memcg, NR_LRU_BASE + lru);
> +               if (tree)
> +                       nr += memcg_page_state(memcg, NR_LRU_BASE + lru);
> +               else
> +                       nr += memcg_page_state_local(memcg, NR_LRU_BASE + lru);
>         }
>         return nr;
>  }
>
> Then we get:
>
>                      text    data     bss     dec     hex filename
> now:               106705   35641    1024  143370   2300a mm/memcontrol.o
> shakeel:           107111   35657    1024  143792   231b0 mm/memcontrol.o
> shakeel+the-above: 106805   35657    1024  143486   2307e mm/memcontrol.o
>
> Which do we prefer?  The 100-byte patch or the 406-byte patch?

I would go with the 100-byte one. The for-loop is just 5 iteration, so
doing a check in each iteration should not be an issue.

Shakeel