On Thu, 8 Apr 2021 at 02:27, Stephen Boyd <sboyd@xxxxxxxxxx> wrote: > > Quoting Dmitry Baryshkov (2021-04-07 15:57:01) > > On Thu, 8 Apr 2021 at 01:41, Stephen Boyd <sboyd@xxxxxxxxxx> 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@xxxxxxxxxxxx> > > > > Cc: Daniel Palmer <daniel@xxxxxxxx> > > > > Signed-off-by: Dmitry Baryshkov <dmitry.baryshkov@xxxxxxxxxx> > > > > --- > > > > > > > > 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
