Re: [RFC 0/4] clk/driver: platform: Fix kfree() of const memory on setting driver_override

[Krzysztof Kozlowski <krzk@kernel.org>]

On Wed, 20 Feb 2019 at 23:01, Stephen Boyd <sboyd@kernel.org> wrote:
>
> Quoting Krzysztof Kozlowski (2019-02-18 03:14:29)
> > On Mon, 18 Feb 2019 at 11:40, Geert Uytterhoeven <geert@linux-m68k.org> wrote:
> > >
> > > Hi Krzysztof,
> > >
> > > On Mon, Feb 18, 2019 at 11:27 AM Krzysztof Kozlowski <krzk@kernel.org> wrote:
> > > > The problem
> > > > ===========
> > > > Several device types (platform, amba, spi etc.) provide a driver_override
> > > > field.  On sysfs store or during device removal, they kfree() the
> > > > existing value.
> > > >
> > > > However the users are unaware of this and set the driver_override like:
> > > >
> > > >         pdev->driver_override = "exynos5-subcmu";
> > > >
> > > > which obviously leads to error.
> > >
> > > IMHO driver_override is not meant to be set by a driver, only from userspace,
> > > for binding the device to vfio (is there another use case?).
> > >
> > > >   clk: samsung: exynos5: Fix kfree() of const memory on setting
> > > >     driver_override
> > > >   slimbus: ngd: Fix kfree() of const memory on setting driver_override
> > >
> > > I see all users set override immediately after allocating a platform device.
> > > Can't they just allocate a platform device using the override name instead?
> > > What am I missing?
> >
> > For the clk-exynos5-subcmu.c case, driver registers several
> > sub-devices. Each sub-devices is a clock controller associated with
> > power domain. I guess the author wanted to have meaningful names of
> > these sub-devices (DISP, CAM etc. like names of power domains),
> > instead of just exynos5-subcmu.1, exynos5-subcmu.2 ...
> >
> > If driver_override should not be used for such case, then I could
> > replace it with what you said.
>
> Looking at the introduction of the code into the platform bus makes it
> sound like it was all for vfio devices. If the clk driver doesn't need
> it for that purpose and can get by with more generic names then it seems
> best to avoid using it entirely. So can you do that and resend the first
> patch in this series too? Effectively splitting the clk parts from the
> larger issue of kfree()ing of const memory.

Sure, let me send just the clk parts.

BR,
Krzysztof