Hello Lukasz,
On Wed, Sep 24, 2014 at 10:27:11AM +0200, Lukasz Majewski wrote:
> Pointer to the uninitialized max_state variable is passed to get the
> maximal cooling state.
> For CPU cooling device (cpu_cooling.c) the cpufreq_get_max_state() is
> called, which even when error occurs will return the max_state variable
> unchanged.
> Since error for ->get_max_state() is not checked, the automatically
Good that you added a fix in your series for this.
> allocated value of max_state is used for (upper > max_state) comparison.
> For any possible max_state value it is very unlikely that it will be less
> than upper.
> As a consequence, the cooling device is bind even without the backed
> cpufreq table initialized.
>
> This initialization will prevent from accidental binding trip points to
> cpu freq cooling frequencies when cpufreq driver itself is not yet fully
> initialized.
Although I agree with the fix, as long as we also include a check for
the .get_max_state return value, I believe the problem you are
describing is about initialization sequence.
In general, I believe we need a better
sequencing between thermal and cpufreq subsystems. One way out is to
include a check for cpufreq driver in the thermal driver, and return
-EPROBE_DEFER when cpufreq is not ready.
>
> Signed-off-by: Lukasz Majewski <[email protected]>
> ---
> drivers/thermal/thermal_core.c | 2 +-
> 1 file changed, 1 insertion(+), 1 deletion(-)
>
> diff --git a/drivers/thermal/thermal_core.c b/drivers/thermal/thermal_core.c
> index 454884a..747618a 100644
> --- a/drivers/thermal/thermal_core.c
> +++ b/drivers/thermal/thermal_core.c
> @@ -927,7 +927,7 @@ int thermal_zone_bind_cooling_device(struct thermal_zone_device *tz,
> struct thermal_instance *pos;
> struct thermal_zone_device *pos1;
> struct thermal_cooling_device *pos2;
> - unsigned long max_state;
> + unsigned long max_state = 0;
> int result;
>
> if (trip >= tz->trips || (trip < 0 && trip != THERMAL_TRIPS_NONE))
> --
> 2.0.0.rc2
>
Hi Eduardo,
> Hello Lukasz,
>
> On Wed, Sep 24, 2014 at 10:27:11AM +0200, Lukasz Majewski wrote:
> > Pointer to the uninitialized max_state variable is passed to get the
> > maximal cooling state.
> > For CPU cooling device (cpu_cooling.c) the cpufreq_get_max_state()
> > is called, which even when error occurs will return the max_state
> > variable unchanged.
> > Since error for ->get_max_state() is not checked, the automatically
>
>
> Good that you added a fix in your series for this.
>
>
> > allocated value of max_state is used for (upper > max_state)
> > comparison. For any possible max_state value it is very unlikely
> > that it will be less than upper.
> > As a consequence, the cooling device is bind even without the backed
> > cpufreq table initialized.
> >
> > This initialization will prevent from accidental binding trip
> > points to cpu freq cooling frequencies when cpufreq driver itself
> > is not yet fully initialized.
>
> Although I agree with the fix, as long as we also include a check for
> the .get_max_state return value, I believe the problem you are
> describing is about initialization sequence.
As you pointed out - the problem here is with initialization sequence.
Thermal and cpufreq cores are initialized very early.
However, the get_max_state() for cpu_cooling.c device (as it is the
pervasive way of cooling things) accesses cpufreq policy to get the
freq_table and count available states.
The issue here is with late initialization of cpufreq policy.
Up till now the cpu_cooling device was bind even when the
get_max_state() returned -EINVAL and everything worked after late
cpufreq policy initialization. However, during this time window the
thermal driver is not configured.
>
> In general, I believe we need a better
> sequencing between thermal and cpufreq subsystems. One way out is to
> include a check for cpufreq driver in the thermal driver, and return
> -EPROBE_DEFER when cpufreq is not ready.
I think that we could return -EPROBE_DEFER when cpufreq's policy is not
yet available and subscribe to cpufreq notifier to call
bind_cooling_device then.
Let's wait for Zhang opinion since he looks after the thermal core code.
>
> >
> > Signed-off-by: Lukasz Majewski <[email protected]>
> > ---
> > drivers/thermal/thermal_core.c | 2 +-
> > 1 file changed, 1 insertion(+), 1 deletion(-)
> >
> > diff --git a/drivers/thermal/thermal_core.c
> > b/drivers/thermal/thermal_core.c index 454884a..747618a 100644
> > --- a/drivers/thermal/thermal_core.c
> > +++ b/drivers/thermal/thermal_core.c
> > @@ -927,7 +927,7 @@ int thermal_zone_bind_cooling_device(struct
> > thermal_zone_device *tz, struct thermal_instance *pos;
> > struct thermal_zone_device *pos1;
> > struct thermal_cooling_device *pos2;
> > - unsigned long max_state;
> > + unsigned long max_state = 0;
> > int result;
> >
> > if (trip >= tz->trips || (trip < 0 && trip !=
> > THERMAL_TRIPS_NONE)) --
> > 2.0.0.rc2
> >
>
--
Best regards,
Lukasz Majewski
Samsung R&D Institute Poland (SRPOL) | Linux Platform Group
On Fri, Oct 03, 2014 at 10:26:28AM +0200, Lukasz Majewski wrote:
> Hi Eduardo,
>
> > Hello Lukasz,
> >
> > On Wed, Sep 24, 2014 at 10:27:11AM +0200, Lukasz Majewski wrote:
> > > Pointer to the uninitialized max_state variable is passed to get the
> > > maximal cooling state.
> > > For CPU cooling device (cpu_cooling.c) the cpufreq_get_max_state()
> > > is called, which even when error occurs will return the max_state
> > > variable unchanged.
> > > Since error for ->get_max_state() is not checked, the automatically
> >
> >
> > Good that you added a fix in your series for this.
> >
> >
> > > allocated value of max_state is used for (upper > max_state)
> > > comparison. For any possible max_state value it is very unlikely
> > > that it will be less than upper.
> > > As a consequence, the cooling device is bind even without the backed
> > > cpufreq table initialized.
> > >
> > > This initialization will prevent from accidental binding trip
> > > points to cpu freq cooling frequencies when cpufreq driver itself
> > > is not yet fully initialized.
> >
> > Although I agree with the fix, as long as we also include a check for
> > the .get_max_state return value, I believe the problem you are
> > describing is about initialization sequence.
>
> As you pointed out - the problem here is with initialization sequence.
> Thermal and cpufreq cores are initialized very early.
>
> However, the get_max_state() for cpu_cooling.c device (as it is the
> pervasive way of cooling things) accesses cpufreq policy to get the
> freq_table and count available states.
>
> The issue here is with late initialization of cpufreq policy.
>
> Up till now the cpu_cooling device was bind even when the
> get_max_state() returned -EINVAL and everything worked after late
> cpufreq policy initialization. However, during this time window the
> thermal driver is not configured.
>
> >
> > In general, I believe we need a better
> > sequencing between thermal and cpufreq subsystems. One way out is to
> > include a check for cpufreq driver in the thermal driver, and return
> > -EPROBE_DEFER when cpufreq is not ready.
>
> I think that we could return -EPROBE_DEFER when cpufreq's policy is not
> yet available and subscribe to cpufreq notifier to call
> bind_cooling_device then.
I agree with this approach, as long as we keep the existing users of
cpufreq cooling in a working state.
Cheers
>
> Let's wait for Zhang opinion since he looks after the thermal core code.
>
> >
> > >
> > > Signed-off-by: Lukasz Majewski <[email protected]>
> > > ---
> > > drivers/thermal/thermal_core.c | 2 +-
> > > 1 file changed, 1 insertion(+), 1 deletion(-)
> > >
> > > diff --git a/drivers/thermal/thermal_core.c
> > > b/drivers/thermal/thermal_core.c index 454884a..747618a 100644
> > > --- a/drivers/thermal/thermal_core.c
> > > +++ b/drivers/thermal/thermal_core.c
> > > @@ -927,7 +927,7 @@ int thermal_zone_bind_cooling_device(struct
> > > thermal_zone_device *tz, struct thermal_instance *pos;
> > > struct thermal_zone_device *pos1;
> > > struct thermal_cooling_device *pos2;
> > > - unsigned long max_state;
> > > + unsigned long max_state = 0;
> > > int result;
> > >
> > > if (trip >= tz->trips || (trip < 0 && trip !=
> > > THERMAL_TRIPS_NONE)) --
> > > 2.0.0.rc2
> > >
> >
>
>
>
> --
> Best regards,
>
> Lukasz Majewski
>
> Samsung R&D Institute Poland (SRPOL) | Linux Platform Group