Received-SPF: pass (google.com: best guess record for domain of linux-kernel-owner@vger.kernel.org designates 209.132.180.67 as permitted sender) client-ip=209.132.180.67;
MIME-Version: 1.0
References: <20190802083541.12602-1-hslester96@gmail.com> <CA+FuTSc8WBx2PCUhn-sLtYHQR-OROXm2pUN9SDj7P-Bd8432UQ@mail.gmail.com>
 <CANhBUQ2TRr4RuSmjaRYPXHZpVw_-2awXvWNjjdvV_z1yoGdkXA@mail.gmail.com>
In-Reply-To: <CANhBUQ2TRr4RuSmjaRYPXHZpVw_-2awXvWNjjdvV_z1yoGdkXA@mail.gmail.com>
From:   Willem de Bruijn <willemdebruijn.kernel@gmail.com>
Date:   Fri, 2 Aug 2019 10:52:28 -0400
Message-ID: <CAF=yD-+3tzufyOnK4suJnovrhX_=4sPqWOsjOcETGG3cA9+MdA@mail.gmail.com>
Subject: Re: [PATCH v2 1/2] cxgb4: sched: Use refcount_t for refcount
To:     Chuhong Yuan <hslester96@gmail.com>
Cc:     Vishal Kulkarni <vishal@chelsio.com>,
        "David S . Miller" <davem@davemloft.net>,
        Network Development <netdev@vger.kernel.org>,
        linux-kernel <linux-kernel@vger.kernel.org>
Content-Type: text/plain; charset="UTF-8"
Content-Transfer-Encoding: quoted-printable
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk

On Fri, Aug 2, 2019 at 10:27 AM Chuhong Yuan <hslester96@gmail.com> wrote:
>
> Willem de Bruijn <willemdebruijn.kernel@gmail.com> =E4=BA=8E2019=E5=B9=B4=
8=E6=9C=882=E6=97=A5=E5=91=A8=E4=BA=94 =E4=B8=8B=E5=8D=889:40=E5=86=99=E9=
=81=93=EF=BC=9A
> >
> > On Fri, Aug 2, 2019 at 4:36 AM Chuhong Yuan <hslester96@gmail.com> wrot=
e:
> > >
> > > refcount_t is better for reference counters since its
> > > implementation can prevent overflows.
> > > So convert atomic_t ref counters to refcount_t.
> > >
> > > Signed-off-by: Chuhong Yuan <hslester96@gmail.com>
> > > ---
> > > Changes in v2:
> > >   - Convert refcount from 0-base to 1-base.
> >
> > This changes the initial value from 0 to 1, but does not change the
> > release condition. So this introduces an accounting bug?
>
> I have noticed this problem and have checked other files which use refcou=
nt_t.
> I find although the refcounts are 1-based, they still use
> refcount_dec_and_test()
> to check whether the resource should be released.
> One example is drivers/char/mspec.c.
> Therefore I think this is okay and do not change the release condition.

Indeed it is fine to use refcount_t with a model where the initial
allocation already accounts for the first reference and thus
initializes with refcount_set(.., 1).

But it is not correct to just change a previously zero initialization
to one. As now an extra refcount_dec will be needed to release state.
But the rest of the code has not changed, so this extra decrement will
not happen.

For a correct conversion, see for instance commits

  commit db5bce32fbe19f0c7482fb5a40a33178bbe7b11b
  net: prepare (struct ubuf_info)->refcnt conversion

and

  commit c1d1b437816f0afa99202be3cb650c9d174667bc
  net: convert (struct ubuf_info)->refcnt to refcount_t

The second makes a search-and-replace style API change like your
patches (though also notice the additional required #include).

But the other patch is needed first to change both the initial
atomic_set *and* at least one atomic_inc, to maintain the same
reference count over the object's lifetime.

That change requires understanding of the object's lifecycle, so I
suggest only making those changes when aware of that whole data path.