PM / Domains: Fix potential deadlock while adding/removing subdomains
authorUlf Hansson <ulf.hansson@linaro.org>
Wed, 27 Jan 2016 08:10:37 +0000 (09:10 +0100)
committerRafael J. Wysocki <rafael.j.wysocki@intel.com>
Wed, 27 Jan 2016 22:01:25 +0000 (23:01 +0100)
We must preserve the same order of how we acquire and release the lock for
genpd, as otherwise we may encounter deadlocks.

The power on phase of a genpd starts by acquiring its lock. Then it walks
the hierarchy of its parent domains to be able to power on these first, as
per design of genpd.

From a locking perspective this means the locks of the parents becomes
acquired after the lock of the subdomain.

Let's fix pm_genpd_add|remove_subdomain() to maintain the same order of
acquiring/releasing the genpd lock as being applied in the power on/off
sequence.

Signed-off-by: Ulf Hansson <ulf.hansson@linaro.org>
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
drivers/base/power/domain.c

index e02ddf65bc43b224354fc2659b6773711dccf65a..2ff818c2ae6bab2c18072a22a42984f76fa51ddc 100644 (file)
@@ -1340,8 +1340,8 @@ int pm_genpd_add_subdomain(struct generic_pm_domain *genpd,
        if (!link)
                return -ENOMEM;
 
-       mutex_lock(&genpd->lock);
-       mutex_lock_nested(&subdomain->lock, SINGLE_DEPTH_NESTING);
+       mutex_lock(&subdomain->lock);
+       mutex_lock_nested(&genpd->lock, SINGLE_DEPTH_NESTING);
 
        if (genpd->status == GPD_STATE_POWER_OFF
            &&  subdomain->status != GPD_STATE_POWER_OFF) {
@@ -1364,8 +1364,8 @@ int pm_genpd_add_subdomain(struct generic_pm_domain *genpd,
                genpd_sd_counter_inc(genpd);
 
  out:
-       mutex_unlock(&subdomain->lock);
        mutex_unlock(&genpd->lock);
+       mutex_unlock(&subdomain->lock);
        if (ret)
                kfree(link);
        return ret;
@@ -1386,7 +1386,8 @@ int pm_genpd_remove_subdomain(struct generic_pm_domain *genpd,
        if (IS_ERR_OR_NULL(genpd) || IS_ERR_OR_NULL(subdomain))
                return -EINVAL;
 
-       mutex_lock(&genpd->lock);
+       mutex_lock(&subdomain->lock);
+       mutex_lock_nested(&genpd->lock, SINGLE_DEPTH_NESTING);
 
        if (!list_empty(&subdomain->slave_links) || subdomain->device_count) {
                pr_warn("%s: unable to remove subdomain %s\n", genpd->name,
@@ -1399,22 +1400,19 @@ int pm_genpd_remove_subdomain(struct generic_pm_domain *genpd,
                if (link->slave != subdomain)
                        continue;
 
-               mutex_lock_nested(&subdomain->lock, SINGLE_DEPTH_NESTING);
-
                list_del(&link->master_node);
                list_del(&link->slave_node);
                kfree(link);
                if (subdomain->status != GPD_STATE_POWER_OFF)
                        genpd_sd_counter_dec(genpd);
 
-               mutex_unlock(&subdomain->lock);
-
                ret = 0;
                break;
        }
 
 out:
        mutex_unlock(&genpd->lock);
+       mutex_unlock(&subdomain->lock);
 
        return ret;
 }