Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 23.128.96.18 as permitted sender) client-ip=23.128.96.18;
Date:   Tue, 6 Apr 2021 15:43:31 +0530
From:   Bharata B Rao <bharata@linux.ibm.com>
To:     Roman Gushchin <guro@fb.com>
Cc:     Yang Shi <shy828301@gmail.com>,
        Kirill Tkhai <ktkhai@virtuozzo.com>,
        Shakeel Butt <shakeelb@google.com>,
        Johannes Weiner <hannes@cmpxchg.org>,
        Andrew Morton <akpm@linux-foundation.org>,
        Linux Kernel Mailing List <linux-kernel@vger.kernel.org>,
        Linux MM <linux-mm@kvack.org>,
        Linux FS-devel Mailing List <linux-fsdevel@vger.kernel.org>,
        aneesh.kumar@linux.ibm.com
Subject: Re: High kmalloc-32 slab cache consumption with 10k containers
Message-ID: <20210406101331.GB1354243@in.ibm.com>
Reply-To: bharata@linux.ibm.com
References: <20210405054848.GA1077931@in.ibm.com>
 <CAHbLzko-17bUWdxmOi-p2_MLSbsMCvhjKS1ktnBysC5dN_W90A@mail.gmail.com>
 <YGtZNLhXjv8RegTK@carbon.dhcp.thefacebook.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <YGtZNLhXjv8RegTK@carbon.dhcp.thefacebook.com>
Precedence: bulk

On Mon, Apr 05, 2021 at 11:38:44AM -0700, Roman Gushchin wrote:
> > > @@ -534,7 +521,17 @@ static void memcg_drain_list_lru_node(struct list_lru *lru, int nid,
> > >         spin_lock_irq(&nlru->lock);
> > >
> > >         src = list_lru_from_memcg_idx(nlru, src_idx);
> > > +       if (!src)
> > > +               goto out;
> > > +
> > >         dst = list_lru_from_memcg_idx(nlru, dst_idx);
> > > +       if (!dst) {
> > > +               /* TODO: Use __GFP_NOFAIL? */
> > > +               dst = kmalloc(sizeof(struct list_lru_one), GFP_ATOMIC);
> > > +               init_one_lru(dst);
> > > +               memcg_lrus = rcu_dereference_protected(nlru->memcg_lrus, true);
> > > +               memcg_lrus->lru[dst_idx] = dst;
> > > +       }
> 
> Hm, can't we just reuse src as dst in this case?
> We don't need src anymore and we're basically allocating dst to move all data from src.

Yes, we can do that and it would be much simpler.

> If not, we can allocate up to the root memcg every time to avoid having
> !dst case and fiddle with __GFP_NOFAIL.
> 
> Otherwise I like the idea and I think it might reduce the memory overhead
> especially on (very) big machines.

Yes, I will however have to check if the callers of list_lru_add() are capable
of handling failure which can happen with this approach if kmalloc fails.

Regards,
Bharata.