When cpufreq online fails, the policy->cpus mask is not cleared and
policy->rwsem is released too early, so the driver can be invoked
via the cpuinfo_cur_freq sysfs attribute while its ->offline() or
->exit() callbacks are being run.

Take policy->clk as an example:

static int cpufreq_online(unsigned int cpu)
{
  ...
  // policy->cpus != 0 at this time
  down_write(&policy->rwsem);
  ret = cpufreq_add_dev_interface(policy);
  up_write(&policy->rwsem);

  return 0;

out_destroy_policy:
	for_each_cpu(j, policy->real_cpus)
		remove_cpu_dev_symlink(policy, get_cpu_device(j));
    up_write(&policy->rwsem);
...
out_exit_policy:
  if (cpufreq_driver->exit)
    cpufreq_driver->exit(policy);
      clk_put(policy->clk);
      // policy->clk is a wild pointer
...
                                    ^
                                    |
                            Another process access
                            __cpufreq_get
                              cpufreq_verify_current_freq
                                cpufreq_generic_get
                                  // acces wild pointer of policy->clk;
                                    |
                                    |
out_offline_policy:                 |
  cpufreq_policy_free(policy);      |
    // deleted here, and will wait for no body reference
    cpufreq_policy_put_kobj(policy);
}

Address this by modifying cpufreq_online() to release policy->rwsem
in the error path after the driver callbacks have run and to clear
policy->cpus before releasing the semaphore.

Fixes: 7106e02b ("cpufreq: release policy->rwsem on error")
Signed-off-by: Schspa Shi <schspa@gmail.com>
[ rjw: Subject and changelog edits ]
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>