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: <20200224094158.28761-1-brgl@bgdev.pl> <20200224094158.28761-3-brgl@bgdev.pl>
 <CACRpkdZSooH+mXbimgT-hnaC2gO1nTi+rY7UmUhVg9bk1j+Eow@mail.gmail.com>
In-Reply-To: <CACRpkdZSooH+mXbimgT-hnaC2gO1nTi+rY7UmUhVg9bk1j+Eow@mail.gmail.com>
From:   Bartosz Golaszewski <brgl@bgdev.pl>
Date:   Fri, 13 Mar 2020 15:47:28 +0100
Message-ID: <CAMRc=Mf2Mx+rB7du8D66WP=Js0wuK8x44aT9H2q6JhLJvrOcVQ@mail.gmail.com>
Subject: Re: [PATCH 2/3] gpiolib: use kref in gpio_desc
To:     Linus Walleij <linus.walleij@linaro.org>
Cc:     Srinivas Kandagatla <srinivas.kandagatla@linaro.org>,
        Khouloud Touil <ktouil@baylibre.com>,
        Geert Uytterhoeven <geert@linux-m68k.org>,
        "open list:GPIO SUBSYSTEM" <linux-gpio@vger.kernel.org>,
        "linux-kernel@vger.kernel.org" <linux-kernel@vger.kernel.org>,
        Bartosz Golaszewski <bgolaszewski@baylibre.com>
Content-Type: text/plain; charset="UTF-8"
Content-Transfer-Encoding: quoted-printable
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk

czw., 12 mar 2020 o 11:35 Linus Walleij <linus.walleij@linaro.org> napisa=
=C5=82(a):
>
> Hi Bartosz,
>
> I'm struggling to figure out if this is the right way to count
> references for gpio descriptors.
>
> I cleared up the situation of why we don't want to add kref
> to gpio_chip in the previous message: I think we got that covered.
> (If I'm not wrong about it, and I am frequently wrong.)
>
> This mail is about contrasting the suggested gpio_desc
> kref with the existing managed resources, i.e. the
> devm_* mechanisms.
>
> devm_* macros are elusive because they do not use
> reference counting at all.
>
> Instead they put every devm_* requested resource with
> a destruction function on a list associated with the struct
> device. Functions get put on that list when we probe a
> device driver, and the list is iterated and all release functions
> are called when we exit .probe() with error or after calling the
> optional .remove() function on the module. (More or less.)
>
> This means anything devm_* managed lives and dies
> with the device driver attaching to the device.
> Documentation/driver-api/driver-model/devres.rst
>
> If the intention of the patch is that this action is associated
> with the detachment of the driver, then we are reinventing
> the wheel we already invented.
>

In this case I was thinking about a situation where we pass a
requested descriptor to some other framework (nvmem in this case)
which internally doesn't know anything about who manages this resource
externally. Now we can of course simply not do anything about it and
expect the user (who passed us the descriptor) to handle the resources
correctly. But what happens if the user releases the descriptor not on
driver detach but somewhere else for whatever reason while nvmem
doesn't know about it? It may try to use the descriptor which will now
be invalid. Reference counting in this case would help IMHO.

Bart

> E.g. to devm_* it doesn't really matter if someone else is
> using a struct gpio_desc, or not, but if the current driver
> is using it, it will be kept around until that driver detaches.
> No reference counting needed for that.
>
> So is this related to your problem or do I just get things
> wrong?