On Fri, Sep 11, 2020 at 10:55 PM Shakeel Butt <shakeelb@xxxxxxxxxx> wrote:
>
> On Thu, Sep 10, 2020 at 8:52 PM Muchun Song <songmuchun@xxxxxxxxxxxxx> wrote:
> >
> > On Fri, Sep 11, 2020 at 12:02 AM Shakeel Butt <shakeelb@xxxxxxxxxx> wrote:
> > >
> > > On Thu, Sep 10, 2020 at 1:46 AM Muchun Song <songmuchun@xxxxxxxxxxxxx> wrote:
> > > >
> > > > In the cgroup v1, we have a numa_stat interface. This is useful for
> > > > providing visibility into the numa locality information within an
> > > > memcg since the pages are allowed to be allocated from any physical
> > > > node. One of the use cases is evaluating application performance by
> > > > combining this information with the application's CPU allocation.
> > > > But the cgroup v2 does not. So this patch adds the missing information.
> > > >
> > > > Signed-off-by: Muchun Song <songmuchun@xxxxxxxxxxxxx>
> > > > ---
> > >
> > > I am actually working on exposing this info on v2 as well.
> > >
> > > > mm/memcontrol.c | 46 ++++++++++++++++++++++++++++++++++++++++++++--
> > > > 1 file changed, 44 insertions(+), 2 deletions(-)
> > > >
> > > > diff --git a/mm/memcontrol.c b/mm/memcontrol.c
> > > > index 75cd1a1e66c8..c779673f29b2 100644
> > > > --- a/mm/memcontrol.c
> > > > +++ b/mm/memcontrol.c
> > > > @@ -1492,10 +1492,34 @@ static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
> > > > return false;
> > > > }
> > > >
> > > > +#ifdef CONFIG_NUMA
> > > > +static unsigned long memcg_node_page_state(struct mem_cgroup *memcg,
> > > > + unsigned int nid,
> > > > + enum node_stat_item idx)
> > > > +{
> > > > + long x;
> > > > + struct mem_cgroup_per_node *pn;
> > > > + struct lruvec *lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(nid));
> > > > +
> > > > + VM_BUG_ON(nid >= nr_node_ids);
> > > > +
> > > > + pn = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
> > > > + x = atomic_long_read(&pn->lruvec_stat[idx]);
> > > > +#ifdef CONFIG_SMP
> > > > + if (x < 0)
> > > > + x = 0;
> > > > +#endif
> > > > + return x;
> > > > +}
> > > > +#endif
> > > > +
> > > > static char *memory_stat_format(struct mem_cgroup *memcg)
> > > > {
> > > > struct seq_buf s;
> > > > int i;
> > > > +#ifdef CONFIG_NUMA
> > > > + int nid;
> > > > +#endif
> > > >
> > > > seq_buf_init(&s, kmalloc(PAGE_SIZE, GFP_KERNEL), PAGE_SIZE);
> > > > if (!s.buffer)
> > > > @@ -1512,12 +1536,30 @@ static char *memory_stat_format(struct mem_cgroup *memcg)
> > > > * Current memory state:
> > > > */
> > > >
> > >
> > > Let's not break the parsers of memory.stat. I would prefer a separate
> > > interface like v1 i.e. memory.numa_stat.
> >
> > It is also a good idea to expose a new interface like memory.numa_stat.
> >
> > >
> > > > - seq_buf_printf(&s, "anon %llu\n",
> > > > + seq_buf_printf(&s, "anon %llu",
> > > > (u64)memcg_page_state(memcg, NR_ANON_MAPPED) *
> > > > PAGE_SIZE);
> > > > - seq_buf_printf(&s, "file %llu\n",
> > > > +#ifdef CONFIG_NUMA
> > > > + for_each_node_state(nid, N_MEMORY)
> > > > + seq_buf_printf(&s, " N%d=%llu", nid,
> > > > + (u64)memcg_node_page_state(memcg, nid,
> > > > + NR_ANON_MAPPED) *
> > > > + PAGE_SIZE);
> > > > +#endif
> > > > + seq_buf_putc(&s, '\n');
> > > > +
> > > > + seq_buf_printf(&s, "file %llu",
> > > > (u64)memcg_page_state(memcg, NR_FILE_PAGES) *
> > > > PAGE_SIZE);
> > > > +#ifdef CONFIG_NUMA
> > > > + for_each_node_state(nid, N_MEMORY)
> > > > + seq_buf_printf(&s, " N%d=%llu", nid,
> > > > + (u64)memcg_node_page_state(memcg, nid,
> > > > + NR_FILE_PAGES) *
> > > > + PAGE_SIZE);
> > > > +#endif
> > > > + seq_buf_putc(&s, '\n');
> > > > +
> > >
> > > The v1's numa_stat exposes the LRUs, why NR_ANON_MAPPED and NR_FILE_PAGES?
> >
> > If we want to expose the anon per node, we need to add inactive anon and
> > active anon together. Why not use NR_ANON_MAPPED directly?
> >
>
> Active anon plus inactive anon is not equal to NR_ANON_MAPPED. The
> shmem related memory is on anon LRUs but not accounted in
> NR_ANON_MAPPED.
>
> Similarly file LRU can contain MADV_FREE pages which are not accounted
> in NR_FILE_PAGES.
I got it, thanks. Because the "state" interface exposes the anon and
file information. So I think that we also should expose the anon and
file for "numa_stat" per node instead of the lru statistics. Maybe it is
better that we expose both of all the information.
>
> > >
> > > Also I think exposing slab_[un]reclaimable per node would be beneficial as well.
> >
> > Yeah, I agree with you. Maybe kernel_stack and percpu also should
> > be exposed.
> >
> > >
> > > > seq_buf_printf(&s, "kernel_stack %llu\n",
> > > > (u64)memcg_page_state(memcg, NR_KERNEL_STACK_KB) *
> > > > 1024);
> > > > --
> > > > 2.20.1
> > > >
> >
> >
> >
> > --
> > Yours,
> > Muchun
--
Yours,
Muchun
