Re: [PATCH v4 01/11] firmware: raspberrypi: Keep count of all consumers

From: Nicolas Saenz Julienne <nsaenzjulienne@suse.de>
To: Dmitry Torokhov <dmitry.torokhov@gmail.com>,
	Andy Shevchenko <andy.shevchenko@gmail.com>
Cc: "Uwe Kleine-König" <u.kleine-koenig@pengutronix.de>,
	"Linux Kernel Mailing List" <linux-kernel@vger.kernel.org>,
	"Florian Fainelli" <f.fainelli@gmail.com>,
	"Ray Jui" <rjui@broadcom.com>,
	"Scott Branden" <sbranden@broadcom.com>,
	bcm-kernel-feedback-list <bcm-kernel-feedback-list@broadcom.com>,
	linux-pwm@vger.kernel.org,
	"linux-arm Mailing List" <linux-arm-kernel@lists.infradead.org>,
	devicetree <devicetree@vger.kernel.org>,
	"Stefan Wahren" <wahrenst@gmx.net>,
	linux-input <linux-input@vger.kernel.org>,
	"Greg Kroah-Hartman" <gregkh@linuxfoundation.org>,
	"open list:STAGING SUBSYSTEM" <devel@driverdev.osuosl.org>,
	"Philipp Zabel" <p.zabel@pengutronix.de>,
	"open list:GPIO SUBSYSTEM" <linux-gpio@vger.kernel.org>,
	"Linus Walleij" <linus.walleij@linaro.org>,
	linux-clk <linux-clk@vger.kernel.org>,
	"Stephen Boyd" <sboyd@kernel.org>,
	linux-rpi-kernel <linux-rpi-kernel@lists.infradead.org>,
	"Bartosz Golaszewski" <bgolaszewski@baylibre.com>
Subject: Re: [PATCH v4 01/11] firmware: raspberrypi: Keep count of all consumers
Date: Mon, 23 Nov 2020 18:19:35 +0100	[thread overview]
Message-ID: <ac38b89133f80f82b857ad2b4e421b501c2f4826.camel@suse.de> (raw)
In-Reply-To: <20201113072615.GE356503@dtor-ws>

[-- Attachment #1: Type: text/plain, Size: 2548 bytes --]

On Thu, 2020-11-12 at 23:26 -0800, Dmitry Torokhov wrote:
> On Thu, Nov 12, 2020 at 07:52:14PM +0200, Andy Shevchenko wrote:
> > On Thu, Nov 12, 2020 at 6:40 PM Nicolas Saenz Julienne
> > <nsaenzjulienne@suse.de> wrote:
> > > 
> > > When unbinding the firmware device we need to make sure it has no
> > > consumers left. Otherwise we'd leave them with a firmware handle
> > > pointing at freed memory.
> > > 
> > > Keep a reference count of all consumers and introduce rpi_firmware_put()
> > > which will permit automatically decrease the reference count upon
> > > unbinding consumer drivers.
> > 
> > ...
> > 
> > >  /**
> > > - * rpi_firmware_get - Get pointer to rpi_firmware structure.
> > >   * @firmware_node:    Pointer to the firmware Device Tree node.
> > >   *
> > > + * The reference to rpi_firmware has to be released with rpi_firmware_put().
> > > + *
> > >   * Returns NULL is the firmware device is not ready.
> > >   */
> > >  struct rpi_firmware *rpi_firmware_get(struct device_node *firmware_node)
> > >  {
> > >         struct platform_device *pdev = of_find_device_by_node(firmware_node);
> > > +       struct rpi_firmware *fw;
> > > 
> > >         if (!pdev)
> > >                 return NULL;
> > > 
> > > -       return platform_get_drvdata(pdev);
> > > +       fw = platform_get_drvdata(pdev);
> > > +       if (!fw)
> > > +               return NULL;
> > > +
> > > +       if (!kref_get_unless_zero(&fw->consumers))
> > > +               return NULL;
> > 

Hi Andy, Dimitry,

> > Don't we have a more traditional way of doing this, i.e.
> > try_module_get() coupled with get_device() ?
> 
> get_device() will make sure that device is there, but gives no
> assurances that device is bound to a driver, so it will not help with
> the racy access to firmware via platform_get_drvdata() call.

I also looked at using get/put_device() just as a means for refcounting (i.e.
replacing fw->consumers), but I can't make it work either. I'd need a way to
hook up into one of the struct device_ktype release() functions. AFAIK it's not
possible for private uses like this.

IIUC the way to do this would be to bypass platform device and create a special
device class/bus for RPi's firmware dependent devices (I could pretty much copy
SCMI's implementation), but I fear that's overkill.

So, for now I'll stick with the kref based implementation, I'll be happy to
change it if you find a better solution. :)

Regards,
Nicolas

[-- Attachment #2: This is a digitally signed message part --]
[-- Type: application/pgp-signature, Size: 488 bytes --]