| Message ID | 20210406230606.3007138-1-dmitry.baryshkov@linaro.org |
|---|---|
| State | Accepted |
| Commit | 50ce6826a48f119baf2794fa384a64efe9bd84a5 |
| Headers | show |
| Series | clk: fixed: fix double free in resource managed fixed-factor clock | expand |
Hi Dmitry, On Wed, 7 Apr 2021 at 08:06, Dmitry Baryshkov <dmitry.baryshkov@linaro.org> wrote: > > devm_clk_hw_register_fixed_factor_release(), the release function for > the devm_clk_hw_register_fixed_factor(), calls > clk_hw_unregister_fixed_factor(), which will kfree() the clock. However > after that the devres functions will also kfree the allocated data, > resulting in double free/memory corruption. Just call > clk_hw_unregister() instead, leaving kfree() to devres code. Doh. Sorry for not spotting this when I wrote the patch. Thank you for cleaning up after me. Cheers, Daniel
Quoting Dmitry Baryshkov (2021-04-06 16:06:06) > devm_clk_hw_register_fixed_factor_release(), the release function for > the devm_clk_hw_register_fixed_factor(), calls > clk_hw_unregister_fixed_factor(), which will kfree() the clock. However > after that the devres functions will also kfree the allocated data, > resulting in double free/memory corruption. Just call > clk_hw_unregister() instead, leaving kfree() to devres code. > > Reported-by: Rob Clark <robdclark@chromium.org> > Cc: Daniel Palmer <daniel@0x0f.com> > Signed-off-by: Dmitry Baryshkov <dmitry.baryshkov@linaro.org> > --- > > Stephen, this fix affects the DSI PHY rework. Do we have a chance of > getting it into 5.12, otherwise there will be a cross-dependency between > msm-next and clk-next. Think I can get this into the last fixes PR. One question though, I think this follows the pattern that things like clk-divider.c use for devm. Are those also broken?
Quoting Dmitry Baryshkov (2021-04-06 16:06:06) > devm_clk_hw_register_fixed_factor_release(), the release function for > the devm_clk_hw_register_fixed_factor(), calls > clk_hw_unregister_fixed_factor(), which will kfree() the clock. However > after that the devres functions will also kfree the allocated data, > resulting in double free/memory corruption. Just call > clk_hw_unregister() instead, leaving kfree() to devres code. > > Reported-by: Rob Clark <robdclark@chromium.org> > Cc: Daniel Palmer <daniel@0x0f.com> > Signed-off-by: Dmitry Baryshkov <dmitry.baryshkov@linaro.org> > --- Applied to clk-fixes. I also squashed this in to cleanup that ugly cast. ---8<---- diff --git a/drivers/clk/clk-fixed-factor.c b/drivers/clk/clk-fixed-factor.c index 390c16f321a6..4e4b6d367612 100644 --- a/drivers/clk/clk-fixed-factor.c +++ b/drivers/clk/clk-fixed-factor.c @@ -66,12 +66,14 @@ EXPORT_SYMBOL_GPL(clk_fixed_factor_ops); static void devm_clk_hw_register_fixed_factor_release(struct device *dev, void *res) { + struct clk_fixed_factor *fix = res; + /* * We can not use clk_hw_unregister_fixed_factor, since it will kfree() * the hw, resulting in double free. Just unregister the hw and let * devres code kfree() it. */ - clk_hw_unregister(&((struct clk_fixed_factor *)res)->hw); + clk_hw_unregister(&fix->hw); } static struct clk_hw *
On Thu, 8 Apr 2021 at 02:27, Stephen Boyd <sboyd@kernel.org> wrote: > > Quoting Dmitry Baryshkov (2021-04-07 15:57:01) > > On Thu, 8 Apr 2021 at 01:41, Stephen Boyd <sboyd@kernel.org> wrote: > > > > > > Quoting Dmitry Baryshkov (2021-04-06 16:06:06) > > > > devm_clk_hw_register_fixed_factor_release(), the release function for > > > > the devm_clk_hw_register_fixed_factor(), calls > > > > clk_hw_unregister_fixed_factor(), which will kfree() the clock. However > > > > after that the devres functions will also kfree the allocated data, > > > > resulting in double free/memory corruption. Just call > > > > clk_hw_unregister() instead, leaving kfree() to devres code. > > > > > > > > Reported-by: Rob Clark <robdclark@chromium.org> > > > > Cc: Daniel Palmer <daniel@0x0f.com> > > > > Signed-off-by: Dmitry Baryshkov <dmitry.baryshkov@linaro.org> > > > > --- > > > > > > > > Stephen, this fix affects the DSI PHY rework. Do we have a chance of > > > > getting it into 5.12, otherwise there will be a cross-dependency between > > > > msm-next and clk-next. > > > > > > Think I can get this into the last fixes PR. One question though, I > > > think this follows the pattern that things like clk-divider.c use for > > > devm. Are those also broken? > > > > It looks so. See e.g. the devres_release() function. It calls > > (*release) callback, then it will kfree the resource. > > Also see Documentation/driver-api/driver-model/devres.rst, which does > > not kfree() in release functions. > > > > Do you wish for me to send all the fixes? > > > > Yes please send more fixes. They're not high priority though so I'll > probably leave them to bake in next for a week or so. Short story: no other patches needed. Long story: I've checked the rest of devres allocations in clk subsystem. Interesting, they use a bit different pattern: they devres_alloc a pointer to the clock, then they fill the pointer with the new clock data. The release callback would (correctly) free the clock pointer by the devres and then devres code would kfree the devres data (clock pointer). The fixed-factor is unique in this area, because it devres_alloc's a clock instance (rather than the pointer) and then fills it, so it should not be freed in the release callback (only unregistered) with the devres code kfreeing() the instance itself. -- With best wishes Dmitry
Quoting Dmitry Baryshkov (2021-04-07 17:52:01) > > Short story: no other patches needed. > > Long story: > I've checked the rest of devres allocations in clk subsystem. > Interesting, they use a bit different pattern: they devres_alloc a > pointer to the clock, then they fill the pointer with the new clock > data. The release callback would (correctly) free the clock pointer by > the devres and then devres code would kfree the devres data (clock > pointer). > > The fixed-factor is unique in this area, because it devres_alloc's a > clock instance (rather than the pointer) and then fills it, so it > should not be freed in the release callback (only unregistered) with > the devres code kfreeing() the instance itself. > > Cool thanks for checking. Maybe we should change those other callers in clk directory to devres alloc the container structure instead of a pointer. Then we avoid the double allocation. Glad it's not a critical fix though.
diff --git a/drivers/clk/clk-fixed-factor.c b/drivers/clk/clk-fixed-factor.c index 4f7bf3929d6d..390c16f321a6 100644 --- a/drivers/clk/clk-fixed-factor.c +++ b/drivers/clk/clk-fixed-factor.c @@ -66,7 +66,12 @@ EXPORT_SYMBOL_GPL(clk_fixed_factor_ops); static void devm_clk_hw_register_fixed_factor_release(struct device *dev, void *res) { - clk_hw_unregister_fixed_factor(&((struct clk_fixed_factor *)res)->hw); + /* + * We can not use clk_hw_unregister_fixed_factor, since it will kfree() + * the hw, resulting in double free. Just unregister the hw and let + * devres code kfree() it. + */ + clk_hw_unregister(&((struct clk_fixed_factor *)res)->hw); } static struct clk_hw *
devm_clk_hw_register_fixed_factor_release(), the release function for the devm_clk_hw_register_fixed_factor(), calls clk_hw_unregister_fixed_factor(), which will kfree() the clock. However after that the devres functions will also kfree the allocated data, resulting in double free/memory corruption. Just call clk_hw_unregister() instead, leaving kfree() to devres code. Reported-by: Rob Clark <robdclark@chromium.org> Cc: Daniel Palmer <daniel@0x0f.com> Signed-off-by: Dmitry Baryshkov <dmitry.baryshkov@linaro.org> --- Stephen, this fix affects the DSI PHY rework. Do we have a chance of getting it into 5.12, otherwise there will be a cross-dependency between msm-next and clk-next. --- drivers/clk/clk-fixed-factor.c | 7 ++++++- 1 file changed, 6 insertions(+), 1 deletion(-)