Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 23.128.96.37 as permitted sender) client-ip=23.128.96.37;
Content-Type: text/plain;
        charset=utf-8
Mime-Version: 1.0
Subject: Re: [PATCH v4 2/5] zswap: make shrinking memcg-aware
From:   Muchun Song <muchun.song@linux.dev>
In-Reply-To: <20231101012614.186996-1-nphamcs@gmail.com>
Date:   Wed, 1 Nov 2023 11:06:26 +0800
Cc:     Andrew Morton <akpm@linux-foundation.org>,
        Johannes Weiner <hannes@cmpxchg.org>,
        cerasuolodomenico@gmail.com, Yosry Ahmed <yosryahmed@google.com>,
        sjenning@redhat.com, ddstreet@ieee.org, vitaly.wool@konsulko.com,
        Michal Hocko <mhocko@kernel.org>,
        Roman Gushchin <roman.gushchin@linux.dev>,
        Shakeel Butt <shakeelb@google.com>,
        Chris Li <chrisl@kernel.org>, Linux-MM <linux-mm@kvack.org>,
        kernel-team@meta.com, LKML <linux-kernel@vger.kernel.org>
Content-Transfer-Encoding: quoted-printable
Message-Id: <B0D87683-DD13-4787-8188-98CEBF561B99@linux.dev>
References: <20231024203302.1920362-3-nphamcs@gmail.com>
 <20231101012614.186996-1-nphamcs@gmail.com>
To:     Nhat Pham <nphamcs@gmail.com>
Precedence: bulk


> On Nov 1, 2023, at 09:26, Nhat Pham <nphamcs@gmail.com> wrote:
>=20
> cc-ing Johannes, Roman, Shakeel, Muchun since you all know much more
> about memory controller + list_lru reparenting logic than me.
>=20
> There seems to be a race between memcg offlining and zswap=E2=80=99s
> cgroup-aware LRU implementation:
>=20
> CPU0                            CPU1
> zswap_lru_add()                 mem_cgroup_css_offline()
>    get_mem_cgroup_from_objcg()
>                                    memcg_offline_kmem()
>                                        memcg_reparent_objcgs()
>                                        memcg_reparent_list_lrus()
>                                            memcg_reparent_list_lru()
>                                                =
memcg_reparent_list_lru_node()
>    list_lru_add()
>                                                memcg_list_lru_free()
>=20
>=20
> Essentially: on CPU0, zswap gets the memcg from the entry's objcg
> (before the objcgs are reparented). Then it performs list_lru_add()
> after the list_lru entries reparenting =
(memcg_reparent_list_lru_node())
> step. If the list_lru of the memcg being offlined has not been freed
> (i.e before the memcg_list_lru_free() call), then the list_lru_add()
> call would succeed - but the list will be freed soon after. The new

No worries.  list_lru_add() will add the object to the lru list of
the parent of the memcg being offlined, because the ->kmemcg_id of the
memcg being offlined will be changed to its parent's ->kmemcg_id before =
memcg_reparent_list_lru().

Muchun,
Thanks

> zswap entry as a result will not be subjected to future reclaim
> attempt. IOW, this list_lru_add() call is effectively swallowed. And
> worse, there might be a crash when we invalidate the zswap_entry in =
the
> future (which will perform a list_lru removal).
>=20
> Within get_mem_cgroup_from_objcg(), none of the following seem
> sufficient to prevent this race:
>=20
>    1. Perform the objcg-to-memcg lookup inside a rcu_read_lock()
>    section.
>    2. Checking if the memcg is freed yet (with css_tryget()) (what
>    we're currently doing in this patch series).
>    3. Checking if the memcg is still online (with css_tryget_online())
>    The memcg can still be offlined down the line.
>=20
>=20
> I've discussed this privately with Johannes, and it seems like the
> cleanest solution here is to move the reparenting logic down to =
release
> stage. That way, when get_mem_cgroup_from_objcg() returns,
> zswap_lru_add() is given an memcg that is reparenting-safe (until we
> drop the obtained reference).