Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::1:20 as permitted sender) client-ip=2620:137:e000::1:20;
Date:   Mon, 10 Oct 2022 10:22:34 +0200
From:   Michal =?iso-8859-1?Q?Koutn=FD?= <mkoutny@suse.com>
To:     Tejun Heo <tj@kernel.org>
Cc:     cgroups@vger.kernel.org, linux-kernel@vger.kernel.org,
        Zefan Li <lizefan.x@bytedance.com>,
        Johannes Weiner <hannes@cmpxchg.org>,
        Dan Carpenter <dan.carpenter@oracle.com>
Subject: Re: [PATCH v2] cgroup: Reorganize css_set_lock and kernfs path
 processing
Message-ID: <Y0PWSlNmWT+1mkvB@blackbook>
References: <20220905170944.23071-1-mkoutny@suse.com>
 <Yxd/sUQ/NB3NlC6f@slm.duckdns.org>
 <YzQw/EvH9Sb58Au2@blackbook>
 <Yz21I9UpXafWMU0K@slm.duckdns.org>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <Yz21I9UpXafWMU0K@slm.duckdns.org>
Precedence: bulk

On Wed, Oct 05, 2022 at 06:47:31AM -1000, Tejun Heo <tj@kernel.org> wrote:
> Hmm... isn't current's root cgrp guaranteed to be alive?

True on the default hierarchy. v1 hierarchies (singular ones with root
cgroup only) can be unmounted.

> How would cgroup_get_live() fail?

kill_sb is not synchronized via css_set_lock.

> Also, shouldn't cgroup_get() enough for path walking?

If ref count dropped to zero, release callback (css_release_work_fn)
would be queued, cgroup_get would increase the refcount but it won't
cancel this.

Note these were concerns with the first version of the patch that also touched
cgroup_show_path() (that processes v1 hierarchies too). With the
reduction I avoided this.

Strictly speaking, even css_set_lock is unnecessary around
current_cgns_cgroup_from_root() when called with cgrp_dfl_root as the
cset->cgrp_links is not traversed at all.

> If you really wanna do it this way, can you please add a detailed comment
> here why this is safe? But I'd prefer just doing a strightforward ref
> inc/dec around it.

I see the the extraction under css_set_lock without inc/dec turns out
confusing. Let me expand the idea above and avoid css_set_lock
completely (another message).

Michal