Re: [PATCH 1/5] clk: renesas: rzv2h: Fix use-after-free in MSTOP refcount handling

[Lad, Prabhakar]

Hi Geert,

On Fri, Dec 20, 2024 at 8:42 AM Geert Uytterhoeven <geert@xxxxxxxxxxxxxx> wrote:
>
> Hi Prabhakar,
>
> On Fri, Dec 20, 2024 at 9:24 AM Lad, Prabhakar
> <prabhakar.csengg@xxxxxxxxx> wrote:
> > On Thu, Dec 19, 2024 at 4:20 PM Geert Uytterhoeven <geert@xxxxxxxxxxxxxx> wrote:
> > > On Wed, Dec 18, 2024 at 3:20 PM Prabhakar <prabhakar.csengg@xxxxxxxxx> wrote:
> > > > From: Lad Prabhakar <prabhakar.mahadev-lad.rj@xxxxxxxxxxxxxx>
> > > >
> > > > Avoid triggering a `refcount_t: addition on 0; use-after-free.` warning
> > > > when registering a module clock with the same MSTOP configuration. The
> > > > issue arises when a module clock is registered but not enabled, resulting
> > > > in a `ref_cnt` of 0. Subsequent calls to `refcount_inc()` on such clocks
> > > > cause the kernel to warn about use-after-free.
> > > >
> > > > [    0.113529] ------------[ cut here ]------------
> > > > [    0.113537] refcount_t: addition on 0; use-after-free.
> > > > [    0.113576] WARNING: CPU: 2 PID: 1 at lib/refcount.c:25 refcount_warn_saturate+0x120/0x144
> > >
> > > [...]
> > >
> > > > Resolve this by checking the `ref_cnt` value before calling
> > > > `refcount_inc()`. If `ref_cnt` is 0, reset it to 1 using `refcount_set()`.
> > >
> > > Thanks for your patch!
> > >
> > > > Fixes: 7bd4cb3d6b7c ("clk: renesas: rzv2h: Relocate MSTOP-related macros to the family driver")
> > >
> > > The description (from your [PATCH 2/5]?) does not match the commit.
> > >
> > Ouch!
> >
> > > Fixes: 7bd4cb3d6b7c43f0 ("clk: renesas: rzv2h: Add MSTOP support")
> > >
> > > > Signed-off-by: Lad Prabhakar <prabhakar.mahadev-lad.rj@xxxxxxxxxxxxxx>
> > >
> > > > --- a/drivers/clk/renesas/rzv2h-cpg.c
> > > > +++ b/drivers/clk/renesas/rzv2h-cpg.c
> > > > @@ -565,8 +565,12 @@ static struct rzv2h_mstop
> > > >                         continue;
> > > >
> > > >                 if (BUS_MSTOP(clk->mstop->idx, clk->mstop->mask) == mstop_data) {
> > > > -                       if (rzv2h_mod_clock_is_enabled(&clock->hw))
> > > > -                               refcount_inc(&clk->mstop->ref_cnt);
> > > > +                       if (rzv2h_mod_clock_is_enabled(&clock->hw)) {
> > > > +                               if (refcount_read(&clk->mstop->ref_cnt))
> > > > +                                       refcount_inc(&clk->mstop->ref_cnt);
> > > > +                               else
> > > > +                                       refcount_set(&clk->mstop->ref_cnt, 1);
> > > > +                       }
>
> Or simply
>
>     do refcount_set(&clk->mstop->ref_cnt,
>                     refcount_read(&clk->mstop->ref_cnt) +1);
>
> ?
>
> Still, you risk some janitor replacing that by refcount_inc() regardless...
>
Agreed.

> > > >                         return clk->mstop;
> > > >                 }
> > > >         }
> > >
> > > This makes me wonder if refcount is the right abstraction?
> > >
> > You mean as discussed on irc, refcount per mstop bit instead of groups
> > is not OK too? Do you have any other better approach in mind?
>
> I mean if you need such silly workarounds to do a simple increment, is
> refcount_t the right abstraction, instead of a plain atomic_t?
>
OK, I'll switch to the atomic_t variant. For this I will still rebase
my work on [0]  along with atomic_t per mstop bit. Is that OK?

[0] https://lore.kernel.org/all/CAMuHMdUEkN6Z7p=LspP+npB3xs4ui+D9oGG+Q15kQ-mYiaoK-A@xxxxxxxxxxxxxx/

Cheers,
Prabhakar