Re: [PATCH RFC v1] clk: Fix potential NULL dereference in clk_fetch_parent_index()

[Martin Blumenstingl]

Hi Stephen,

On Fri, Aug 16, 2019 at 1:29 AM Stephen Boyd <sboyd@xxxxxxxxxx> wrote:
>
> Quoting Martin Blumenstingl (2019-08-15 15:31:55)
> > Don't compare the parent clock name with a NULL name in the
> > clk_parent_map. This prevents a kernel crash when passing NULL
> > core->parents[i].name to strcmp().
> >
> > An example which triggered this is a mux clock with four parents when
> > each of them is referenced in the clock driver using
> > clk_parent_data.fw_name and then calling clk_set_parent(clk, 3rd_parent)
> > on this mux.
> > In this case the first parent is also the HW default so
> > core->parents[i].hw is populated when the clock is registered. Calling
> > clk_set_parent(clk, 3rd_parent) will then go through all parents and
> > skip the first parent because it's hw pointer doesn't match. For the
> > second parent no hw pointer is cached yet and clk_core_get(core, 1)
> > returns a non-matching pointer (which is correct because we are comparing
> > the second with the third parent). Comparing the result of
> > clk_core_get(core, 2) with the requested parent gives a match. However
> > we don't reach this point because right after the clk_core_get(core, 1)
> > mismatch the old code tried to !strcmp(parent->name, NULL) (where the
> > second argument is actually core->parents[i].name, but that was never
> > populated by the clock driver).
> >
> > Signed-off-by: Martin Blumenstingl <martin.blumenstingl@xxxxxxxxxxxxxx>
> > ---
> > I have seen the original crash when I was testing an MMC driver which
> > is not upstream yet on v5.3-rc4. I'm not sure whether this fix is
> > "correct" (it fixes the crash for me) or where to point the Fixes tag
> > to, it may be one of:
> > - fc0c209c147f ("clk: Allow parents to be specified without string names")
> > - 1a079560b145 ("clk: Cache core in clk_fetch_parent_index() without names")
> >
> > This is meant to be applied on top of v5.3-rc4.
> >
>
> Ah ok. I thought that strcmp() would ignore NULL arguments, but
> apparently not. I can apply this to clk-fixes.
at least ARM [0] and the generic [1] implementations don't

I did not bisect this so do you have any suggestion for a Fixes tag? I
mentioned two candidates above, but I'm not sure which one to use
just let me know, then I'll resend with the fixes tag so you can take
it through clk-fixes


Martin


[0] https://elixir.bootlin.com/linux/v5.2/source/arch/arm/boot/compressed/string.c#L91
[1] https://elixir.bootlin.com/linux/v5.2/source/lib/string.c#L356