Message-ID: <20200622215800.GA326762@carbon.DHCP.thefacebook.com>
Date:   Mon, 22 Jun 2020 14:58:00 -0700
From:   Roman Gushchin <guro@...com>
To:     Shakeel Butt <shakeelb@...gle.com>
CC:     Andrew Morton <akpm@...ux-foundation.org>,
        Christoph Lameter <cl@...ux.com>,
        Johannes Weiner <hannes@...xchg.org>,
        Michal Hocko <mhocko@...nel.org>,
        Linux MM <linux-mm@...ck.org>,
        Vlastimil Babka <vbabka@...e.cz>,
        Kernel Team <kernel-team@...com>,
        LKML <linux-kernel@...r.kernel.org>
Subject: Re: [PATCH v6 17/19] mm: memcg/slab: use a single set of kmem_caches
 for all allocations

On Mon, Jun 22, 2020 at 02:28:54PM -0700, Shakeel Butt wrote:
> On Mon, Jun 22, 2020 at 2:15 PM Roman Gushchin <guro@...com> wrote:
> >
> > On Mon, Jun 22, 2020 at 02:04:29PM -0700, Shakeel Butt wrote:
> > > On Mon, Jun 22, 2020 at 1:37 PM Roman Gushchin <guro@...com> wrote:
> > > >
> > > > On Mon, Jun 22, 2020 at 12:21:28PM -0700, Shakeel Butt wrote:
> > > > > On Mon, Jun 8, 2020 at 4:07 PM Roman Gushchin <guro@...com> wrote:
> > > > > >
> > > > > > Instead of having two sets of kmem_caches: one for system-wide and
> > > > > > non-accounted allocations and the second one shared by all accounted
> > > > > > allocations, we can use just one.
> > > > > >
> > > > > > The idea is simple: space for obj_cgroup metadata can be allocated
> > > > > > on demand and filled only for accounted allocations.
> > > > > >
> > > > > > It allows to remove a bunch of code which is required to handle
> > > > > > kmem_cache clones for accounted allocations. There is no more need
> > > > > > to create them, accumulate statistics, propagate attributes, etc.
> > > > > > It's a quite significant simplification.
> > > > > >
> > > > > > Also, because the total number of slab_caches is reduced almost twice
> > > > > > (not all kmem_caches have a memcg clone), some additional memory
> > > > > > savings are expected. On my devvm it additionally saves about 3.5%
> > > > > > of slab memory.
> > > > > >
> > > > > > Suggested-by: Johannes Weiner <hannes@...xchg.org>
> > > > > > Signed-off-by: Roman Gushchin <guro@...com>
> > > > > > Reviewed-by: Vlastimil Babka <vbabka@...e.cz>
> > > > > > ---
> > > > > [snip]
> > > > > >  static inline void memcg_slab_post_alloc_hook(struct kmem_cache *s,
> > > > > >                                               struct obj_cgroup *objcg,
> > > > > > -                                             size_t size, void **p)
> > > > > > +                                             gfp_t flags, size_t size,
> > > > > > +                                             void **p)
> > > > > >  {
> > > > > >         struct page *page;
> > > > > >         unsigned long off;
> > > > > >         size_t i;
> > > > > >
> > > > > > +       if (!objcg)
> > > > > > +               return;
> > > > > > +
> > > > > > +       flags &= ~__GFP_ACCOUNT;
> > > > > >         for (i = 0; i < size; i++) {
> > > > > >                 if (likely(p[i])) {
> > > > > >                         page = virt_to_head_page(p[i]);
> > > > > > +
> > > > > > +                       if (!page_has_obj_cgroups(page) &&
> > > > >
> > > > > The page is already linked into the kmem_cache, don't you need
> > > > > synchronization for memcg_alloc_page_obj_cgroups().
> > > >
> > > > Hm, yes, in theory we need it. I guess the reason behind why I've never seen any issues
> > > > here is the SLUB percpu partial list.
> > > >
> > > > So in theory we need something like:
> > > >
> > > > diff --git a/mm/slab.h b/mm/slab.h
> > > > index 0a31600a0f5c..44bf57815816 100644
> > > > --- a/mm/slab.h
> > > > +++ b/mm/slab.h
> > > > @@ -237,7 +237,10 @@ static inline int memcg_alloc_page_obj_cgroups(struct page *page,
> > > >         if (!vec)
> > > >                 return -ENOMEM;
> > > >
> > > > -       page->obj_cgroups = (struct obj_cgroup **) ((unsigned long)vec | 0x1UL);
> > > > +       if (cmpxchg(&page->obj_cgroups, 0,
> > > > +                   (struct obj_cgroup **) ((unsigned long)vec | 0x1UL)))
> > > > +               kfree(vec);
> > > > +
> > > >         return 0;
> > > >  }
> > > >
> > > >
> > > > But I wonder if we might put it under #ifdef CONFIG_SLAB?
> > > > Or any other ideas how to make it less expensive?
> > > >
> > > > > What's the reason to remove this from charge_slab_page()?
> > > >
> > > > Because at charge_slab_page() we don't know if we'll ever need
> > > > page->obj_cgroups. Some caches might have only few or even zero
> > > > accounted objects.
> > > >
> > >
> > > If slab_pre_alloc_hook() returns a non-NULL objcg then we definitely
> > > need page->obj_cgroups.  The charge_slab_page() happens between
> > > slab_pre_alloc_hook() & slab_post_alloc_hook(), so, we should be able
> > > to tell if page->obj_cgroups is needed.
> >
> > Yes, but the opposite is not always true: we can reuse the existing page
> > without allocated page->obj_cgroups. In this case charge_slab_page() is
> > not involved at all.
> >
> 
> Hmm yeah, you are right. I missed that.
> 
> >
> > Or do you mean that we can minimize the amount of required synchronization
> > by allocating some obj_cgroups vectors from charge_slab_page()?
> 
> One optimization would be to always pre-allocate page->obj_cgroups for
> kmem_caches with SLAB_ACCOUNT.

Even this is not completely memory overhead-free, because processes belonging
to the root cgroup and kthreads might allocate from such cache.

Anyway, I think I'll go with cmpxchg() for now and will think about possible
optimizations later. Because the allocation happens only once per the lifetime
of a slab page, and is very unlikely racing with a concurrent one on the same page,
the penalty shouldn't be that big.

Thanks!

Powered by blists - more mailing lists