[PATCH v2] regulator: disable disable_work for good in regulator_unregister()

[Chengyu Liu]

regulator_unregister() stops rdev->disable_work with
flush_work(&dw->work), which is a no-op for a delayed_work pending on
the timer wheel: flush_work() only waits for a queued or running work
item.  A disable deferred by a consumer's regulator_disable_deferred()
therefore survives the provider's unbind, and regulator_disable_work()
fires after device_unregister() has freed the rdev, dereferencing it
via container_of().  Use-after-free.

The in-tree users are ASoC DAPM regulator widgets with a 20 ms
power-down delay (sound/soc/soc-dapm.c): the window is the entire
deferral period, not a race between adjacent instructions.

Use disable_delayed_work_sync(): it dequeues a timer-pending instance
and waits for a running callback, and keeps the work disabled so a
concurrent regulator_disable_deferred() cannot re-arm it while the
teardown proceeds.

Found by static analysis.

Cc: stable@xxxxxxxxxxxxxxx
Signed-off-by: Chengyu Liu <lui2012qwq@xxxxxxxxx>
---
v2: switch to disable_delayed_work_sync() (Mark Brown): also keeps the
    work disabled so nothing can re-arm it past the teardown.

 drivers/regulator/core.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/drivers/regulator/core.c b/drivers/regulator/core.c
index 6a4008f387b5..4ba0457c4747 100644
--- a/drivers/regulator/core.c
+++ b/drivers/regulator/core.c
@@ -6300,7 +6300,7 @@ void regulator_unregister(struct regulator_dev *rdev)
 		regulator_put(rdev->supply);
 	}
 
-	flush_work(&rdev->disable_work.work);
+	disable_delayed_work_sync(&rdev->disable_work);
 
 	mutex_lock(&regulator_list_mutex);
 
-- 
2.25.1