Re: [PATCH for-5.10] spi: gpio: Don't leak SPI master in probe error path

From: Andrey Smirnov <andrew.smirnov@gmail.com>
To: Lukas Wunner <lukas@wunner.de>
Cc: Mark Brown <broonie@kernel.org>,
	linux-spi@vger.kernel.org,
	Navid Emamdoost <navid.emamdoost@gmail.com>,
	Linus Walleij <linus.walleij@linaro.org>
Subject: Re: [PATCH for-5.10] spi: gpio: Don't leak SPI master in probe error path
Date: Mon, 16 Nov 2020 15:59:20 -0800	[thread overview]
Message-ID: <CAHQ1cqHx0Bq+g8qFK_QdjGHgKHyLgev_5EAwo3CAF9knGfNVSg@mail.gmail.com> (raw)
In-Reply-To: <20201116230339.GA25992@wunner.de>

On Mon, Nov 16, 2020 at 3:03 PM Lukas Wunner <lukas@wunner.de> wrote:
>
> On Mon, Nov 16, 2020 at 11:23:43AM -0800, Andrey Smirnov wrote:
> > On Mon, Nov 16, 2020 at 12:44 AM Lukas Wunner <lukas@wunner.de> wrote:
> > > If the call to devm_spi_register_master() fails on probe of the GPIO SPI
> > > driver, the spi_master struct is erroneously not freed:
> > >
> > > After allocating the spi_master, its reference count is 1.  The driver
> > > unconditionally decrements the reference count on unbind using a devm
> > > action.  Before calling devm_spi_register_master(), the driver
> > > unconditionally increments the reference count because on success,
> > > that function will decrement the reference count on unbind.  However on
> > > failure, devm_spi_register_master() does *not* decrement the reference
> > > count, so the spi_master is leaked.
> >
> > Not sure I fully understand this. On failure
> > devm_spi_register_master() will return a negative error code which
> > should result in probe failure and release of devres resource, right?
>
> Yes, but that just decrements the refcount from 2 to 1:
>
>     /* refcount initialized to 1 */
>     master = spi_alloc_master(dev, sizeof(*spi_gpio));
>
>     ...
>
>     /* refcount incremented to 2 */
>     return devm_spi_register_master(&pdev->dev, spi_master_get(master));
>
>     ...
>
>     /* on failure of devm_spi_register_master(), refcount decremented to 1
>        by devres action */
>     spi_gpio_put()
>
>
> > > The issue was introduced by commits 8b797490b4db ("spi: gpio: Make sure
> > > spi_master_put() is called in every error path") and 79567c1a321e ("spi:
> > > gpio: Use devm_spi_register_master()"), which sought to plug leaks
> > > introduced by 9b00bc7b901f ("spi: spi-gpio: Rewrite to use GPIO
> > > descriptors") but missed this remaining leak.
> >
> > That extra spi_master_get() that might be problematic was present in
> > the code before 8b797490b4db ("spi: gpio: Make sure spi_master_put()
> > is called in every error path") and I think was first introduced in
> >
> > https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/drivers?h=v5.9-rc4&id=702a4879ec337463f858c8ab467482cce260bf18
> >
> > Or am I missing something?
>
> The extra spi_master_get() was introduced by 79567c1a321e.
> I don't see it in spi-gpio.c before that commit.
>
> Its quite possible that I missed something myself, nobody's perfect.
> But just from code inspection it seems wrong the way it is right now.
>
> Shout if I failed to explain it properly and I'll try again. :)
>

Before 79567c1a321e extra spi_master_get() was a part of
spi_bitbang_start(). Ah, OK, it had it's own local spi_master_put() on
failure, which I lost when inlinging that function. My bad. Good to
see that the with devm_spi_alloc_master() cleanup path is sane and
easy to read once again.

> Thanks,
>
> Lukas