Re: [PATCH] thermal: fix race condition when updating cooling device

[Javi Merino <javi.merino@arm.com>]

On Thu, Jun 02, 2016 at 03:25:31PM +0100, Michele Di Giorgio wrote:
> When multiple thermal zones are bound to the same cooling device, multiple
> kernel threads may want to update the cooling device state by calling
> thermal_cdev_update(). Having cdev not protected by a mutex can lead to a race
> condition. Consider the following situation with two kernel threads k1 and k2:
> 
> 	    Thread k1				Thread k2
>                                     ||
>                                     ||  call thermal_cdev_update()
>                                     ||      ...
>                                     ||      set_cur_state(cdev, target);
>     call power_actor_set_power()    ||
>         ...                         ||
>         instance->target = state;   ||
>         cdev->updated = false;      ||
>                                     ||      cdev->updated = true;
>                                     ||      // completes execution
>     call thermal_cdev_update()      ||
>         // cdev->updated == true    ||
>         return;                     ||
>                                     \/
>                                     time
> 
> k2 has already looped through the thermal instances looking for the deepest
> cooling device state and is preempted right before setting cdev->updated to
> true. Now, k1 runs, modifies the thermal instance state and sets cdev->updated
> to false. Then, k1 is preempted and k2 continues the execution by setting
> cdev->updated to true, therefore preventing k1 from performing the update.
> Notice that this is not an issue if k2 looks at the instance->target modified by
> k1 "after" it is assigned by k1. In fact, in this case the update will happen
> anyway and k1 can safely return immediately from thermal_cdev_update().
> 
> This may lead to a situation where a thermal governor never updates the cooling
> device. For example, this is the case for the step_wise governor: when calling
> the function thermal_zone_trip_update(), the governor may always get a new state
> equal to the old one (which, however, wasn't notified to the cooling device) and
> will therefore skip the update.
> 
> CC: Zhang Rui <rui.zhang@intel.com>
> CC: Eduardo Valentin <edubezval@gmail.com>
> CC: Peter Feuerer <peter@piie.net>
> Reported-by: Toby Huang <toby.huang@arm.com>
> Signed-off-by: Michele Di Giorgio <michele.digiorgio@arm.com>
> ---
> Protecting only the assignment of cdev->updated with mutexes may look
> suspicious, but it is necessary to guarantee synchronization and avoiding the
> situation described in the commit message.
> 
> There are other two possible solutions.
> 
> Moving the cdev->lock mutex outside thermal_cdev_update() and protect both the
> assignment and the function. This would work, but will probably cause many
> issues when updating all the modules that use thermal_cdev_update().
> 
> The other solution is to make cdev->updated an atomic_t, change the if
> condition to an atomic_cmpxchg and extend the critical section to include the
> call to cdev->ops->set_cur_state().

True.  In any case, the mutex needs to cover set_cur_state() in
thermal_cdev_update().  This fixes the race condition, so I'm happy
with it.

Reviewed-by: Javi Merino <javi.merino@arm.com>

>  drivers/thermal/fair_share.c      |  2 ++
>  drivers/thermal/gov_bang_bang.c   |  2 ++
>  drivers/thermal/power_allocator.c |  2 ++
>  drivers/thermal/step_wise.c       |  2 ++
>  drivers/thermal/thermal_core.c    | 10 +++++++---
>  5 files changed, 15 insertions(+), 3 deletions(-)