[PATCH] regulator: core: Avoid potential deadlock on regulator_unregister

[PATCH] regulator: core: Avoid potential deadlock on regulator_unregister

[Charles Keepax]

Lockdep reports the following issue on my setup:

Possible unsafe locking scenario:

CPU0                    CPU1
----                    ----
lock((work_completion)(&(&rdev->disable_work)->work));
                        lock(regulator_list_mutex);
                        lock((work_completion)(&(&rdev->disable_work)->work));
lock(regulator_list_mutex);

The problem is that regulator_unregister takes the
regulator_list_mutex and then calls flush_work on disable_work. But
regulator_disable_work calls regulator_lock_dependent which will
also take the regulator_list_mutex. Resulting in a deadlock if the
flush_work call actually needs to flush the work.

Fix this issue by moving the flush_work outside of the
regulator_list_mutex. The list mutex is not used to guard the point at
which the delayed work is queued, so its use adds no additional safety.

Fixes: f8702f9e4aa7 ("regulator: core: Use ww_mutex for regulators locking")
Signed-off-by: Charles Keepax <ckeepax@opensource.cirrus.com>
---

This patch follows on from my email the other day [1].

After looking at things in more detail I am fairly confident this is
a good fix. I do still have a slight nagging doubt that something
should be protecting this flush_work from additional works being
queued, and I can't see what that is. But as that is definitely not
the regulator_list_mutex the patch is not making this any more
dangerous. In practice I suspect this is fine as nothing should
really be using a regulator that is about to be unregistered,
or really this delayed work is probably the least of the systems
problems.

Thanks,
Charles

[1] https://lore.kernel.org/lkml/20190403135531.GB81578@ediswmail.ad.cirrus.com/

 drivers/regulator/core.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/drivers/regulator/core.c b/drivers/regulator/core.c
index 38442eebddfc5..186a37675b50b 100644
--- a/drivers/regulator/core.c
+++ b/drivers/regulator/core.c
@@ -5059,10 +5059,11 @@ void regulator_unregister(struct regulator_dev *rdev)
 		regulator_put(rdev->supply);
 	}
 
+	flush_work(&rdev->disable_work.work);
+
 	mutex_lock(&regulator_list_mutex);
 
 	debugfs_remove_recursive(rdev->debugfs);
-	flush_work(&rdev->disable_work.work);
 	WARN_ON(rdev->open_count);
 	regulator_remove_coupling(rdev);
 	unset_regulator_supplies(rdev);
-- 
2.11.0

Re: [PATCH] regulator: core: Avoid potential deadlock on regulator_unregister

[Dmitry Osipenko]

04.04.2019 18:32, Charles Keepax пишет:
> Lockdep reports the following issue on my setup:
> 
> Possible unsafe locking scenario:
> 
> CPU0                    CPU1
> ----                    ----
> lock((work_completion)(&(&rdev->disable_work)->work));
>                         lock(regulator_list_mutex);
>                         lock((work_completion)(&(&rdev->disable_work)->work));
> lock(regulator_list_mutex);
> 
> The problem is that regulator_unregister takes the
> regulator_list_mutex and then calls flush_work on disable_work. But
> regulator_disable_work calls regulator_lock_dependent which will
> also take the regulator_list_mutex. Resulting in a deadlock if the
> flush_work call actually needs to flush the work.
> 
> Fix this issue by moving the flush_work outside of the
> regulator_list_mutex. The list mutex is not used to guard the point at
> which the delayed work is queued, so its use adds no additional safety.
> 
> Fixes: f8702f9e4aa7 ("regulator: core: Use ww_mutex for regulators locking")
> Signed-off-by: Charles Keepax <ckeepax@opensource.cirrus.com>
> ---
> 
> This patch follows on from my email the other day [1].
> 
> After looking at things in more detail I am fairly confident this is
> a good fix. I do still have a slight nagging doubt that something
> should be protecting this flush_work from additional works being
> queued, and I can't see what that is. But as that is definitely not
> the regulator_list_mutex the patch is not making this any more
> dangerous. In practice I suspect this is fine as nothing should
> really be using a regulator that is about to be unregistered,
> or really this delayed work is probably the least of the systems
> problems.

If anything tries to use regulator after invoking the regulator_unregister(), then indeed it's already in trouble. The patch looks good to me, thanks.

Reviewed-by: Dmitry Osipenko <digetx@gmail.com>