On Tue, Jun 19, 2018 at 5:49 PM David Rientjes <rientjes@xxxxxxxxxx> wrote:
>
> On Tue, 19 Jun 2018, Shakeel Butt wrote:
>
> > diff --git a/mm/slub.c b/mm/slub.c
> > index a3b8467c14af..731c02b371ae 100644
> > --- a/mm/slub.c
> > +++ b/mm/slub.c
> > @@ -3673,9 +3673,23 @@ static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
> >
> > bool __kmem_cache_empty(struct kmem_cache *s)
> > {
> > - int node;
> > + int cpu, node;
>
> Nit: wouldn't cpu be unused if CONFIG_SLUB_DEBUG is disabled?
>
I think I didn't get the warning as I didn't use #ifdef.
> > struct kmem_cache_node *n;
> >
> > + /*
> > + * slabs_node will always be 0 for !CONFIG_SLUB_DEBUG. So, manually
> > + * check slabs for all cpus.
> > + */
> > + if (!IS_ENABLED(CONFIG_SLUB_DEBUG)) {
> > + for_each_online_cpu(cpu) {
> > + struct kmem_cache_cpu *c;
> > +
> > + c = per_cpu_ptr(s->cpu_slab, cpu);
> > + if (c->page || slub_percpu_partial(c))
> > + return false;
> > + }
> > + }
> > +
> > for_each_kmem_cache_node(s, node, n)
> > if (n->nr_partial || slabs_node(s, node))
> > return false;
>
> Wouldn't it just be better to allow {inc,dec}_slabs_node() to adjust the
> nr_slabs counter instead of doing the per-cpu iteration on every shutdown?
Yes that is doable as the functions {inc,dec}_slabs_node() are called
in slow path. I can move them out of CONFIG_SLUB_DEBUG. I think the
reason 0f389ec63077 ("slub: No need for per node slab counters if
!SLUB_DEBUG") put them under CONFIG_SLUB_DEBUG is because these
counters were only read through sysfs API which were disabled on
!CONFIG_SLUB_DEBUG. However we have a usecase other than sysfs API.
Please let me know if there is any objection to this conversion. For
large machines I think this is preferable approach.
thanks,
Shakeel
