Re: [RFC PATCH] cpufreq/hotplug: Fix cpu-hotplug cpufreq race conditions

[Preeti U Murthy <preeti@linux.vnet.ibm.com>]

On 06/02/2015 11:09 AM, Viresh Kumar wrote:
> On 02-06-15, 11:01, Preeti U Murthy wrote:
>> How will a policy lock help here at all, when cpus from multiple
>> policies are calling into __cpufreq_governor() ? How will a policy lock
>> serialize their entry into cpufreq_governor_dbs() ?
> 
> So different policies don't really depend on each other. The only
> thing common to them are the governor's sysfs files (only if
> governor-per-policy isn't set, i.e. in your case). Those sysfs files
> and their kernel counterpart variables aren't touched unless all the
> policies have EXITED. All these START/STOP calls touch only the data
> relevant to those policies only.

No, dbs_data is a governor wide data structure and not a policy wide
one, which is manipulated in START/STOP calls for drivers where the
CPUFREQ_HAVE_GOVERNOR_PER_POLICY is not set.

So even if we assume that we hold per-policy locks, the following race
is still present. Assume that we have just two cpus which do not have a
governor-per-policy set.

CPU0                        CPU1

 store*                     store*

lock(policy 1)              lock(policy 2)
cpufreq_set_policy()       cpufreq_set_policy()
EXIT() :
dbs-data->usage_count--

INIT()
dbs_data exists
so return
                            EXIT()
                            dbs_data->usage_count -- = 0
                            kfree(dbs_data)

START()
dereference dbs_data
*NULL dereference*


Regards
Preeti U Murthy