[PATCH] power: supply: core: give each supply's epc_sem its own lockdep class
From: Ryan Brue
Date: Tue Nov 10 2026 - 20:53:08 EST
Since commit 9f7f7a859451 ("power_supply: serialize property access with
unregister"), power_supply_get_property() takes the target supply's
epc_sem for read. Every supply's epc_sem is initialised at the same site,
so they all share one lockdep class.
A driver that reads another supply while holding a lock that its
get_property() callback also takes is then reported as a circular
dependency, even though the two semaphores belong to different supplies.
sc27xx_fgu_get_property() does this with data->lock and the charger, and
a fuel gauge driver in preparation did the same with its charger:
WARNING: possible circular locking dependency detected
kworker/2:1/60 is trying to acquire lock:
(mutex_#2){+.+.}-{4:4}, at: mt6397_battery_get_property+0x50/0x23c
but task is already holding lock:
(&psy->epc_sem){.+.+}-{4:4}, at: power_supply_get_property+0x34/0x80
Lockdep then turns itself off for the rest of the boot.
Register a lockdep key for each supply, so that only a supply's own epc_sem
is ordered against its own locks.
Fixes: 9f7f7a859451 ("power_supply: serialize property access with unregister")
Signed-off-by: Ryan Brue <ryanbrue.dev@xxxxxxxxx>
---
Since v7.4-rc1 power_supply_get_property() takes the target supply's
epc_sem, and every supply's epc_sem shares one lockdep class. A driver
that reads another supply under a lock its own get_property() takes is
then reported as a circular dependency, and lockdep switches itself off.
Tested on an MT8173 tablet (Amazon Fire HD 10, 2017) with PROVE_LOCKING:
the report came at boot with a fuel gauge that reads its charger under
its own mutex, and doesn't come with this patch (debug_locks stays 1).
---
drivers/power/supply/power_supply_core.c | 3 +++
include/linux/power_supply.h | 3 +++
2 files changed, 6 insertions(+)
diff --git a/drivers/power/supply/power_supply_core.c b/drivers/power/supply/power_supply_core.c
index de15314ef954..76bbe8772143 100644
--- a/drivers/power/supply/power_supply_core.c
+++ b/drivers/power/supply/power_supply_core.c
@@ -1662,6 +1662,7 @@ static void power_supply_dev_release(struct device *dev)
struct power_supply *psy = to_power_supply(dev);
dev_dbg(dev, "%s\n", __func__);
+ lockdep_unregister_key(&psy->epc_key);
kfree(psy);
}
@@ -1765,6 +1766,7 @@ __power_supply_register(struct device *parent,
if (!psy)
return ERR_PTR(-ENOMEM);
+ lockdep_register_key(&psy->epc_key);
dev = &psy->dev;
device_initialize(dev);
@@ -1810,6 +1812,7 @@ __power_supply_register(struct device *parent,
spin_lock_init(&psy->changed_lock);
init_rwsem(&psy->epc_sem);
+ lockdep_set_class(&psy->epc_sem, &psy->epc_key);
init_rwsem(&psy->extensions_sem);
INIT_LIST_HEAD(&psy->extensions);
diff --git a/include/linux/power_supply.h b/include/linux/power_supply.h
index cce61137aef4..3cfe2520862e 100644
--- a/include/linux/power_supply.h
+++ b/include/linux/power_supply.h
@@ -371,8 +371,11 @@ struct power_supply {
* the callback is still using it.
* Must not be shared with extensions_sem: callbacks may read their own
* properties, which takes that one for read.
+ * Each supply has its own lockdep class: a driver may read another
+ * supply's properties while holding a lock its own callbacks take.
*/
struct rw_semaphore epc_sem;
+ struct lock_class_key epc_key;
struct rw_semaphore extensions_sem; /* protects "extensions" */
struct list_head extensions;
#ifdef CONFIG_THERMAL
---
base-commit: 6628e7d4c6b25c8d18f1345e3a20d21829aae7e4
change-id: 20261103-power-supply-epc-lockdep-key-2af05a195ccc
Best regards,
--
Ryan Brue <ryanbrue.dev@xxxxxxxxx>