Messages in this thread Patch in this message | | | Date | Tue, 11 Aug 2020 02:07:22 +0200 | From | Michał Mirosław <> | Subject | Re: regulator: deadlock vs memory reclaim |
| |
On Mon, Aug 10, 2020 at 11:56:13PM +0300, Dmitry Osipenko wrote: > 10.08.2020 23:21, Dmitry Osipenko пишет: > > 10.08.2020 23:18, Michał Mirosław пишет: > >> On Mon, Aug 10, 2020 at 11:15:28PM +0300, Dmitry Osipenko wrote: > >>> 10.08.2020 23:09, Michał Mirosław пишет: > >>>> At first I also thought so, but there's more. Below is a lockdep > >>>> complaint with your patch applied. I did a similar patch and then two more > >>>> (following) and that is still not enough (sysfs/debugfs do allocations, > >>>> too). > >>> Then it should be good to move the locking for init_coupling() like I > >>> suggested and use GFP_NOWAIT for the two other cases. It all could be a > >>> single small patch. Could you please check whether GFP_NOWAIT helps? > >> > >> This would be equivalent to my patches. Problem with sysfs and debugfs > >> remains as they don't have the option of GFP_NOWAIT. This needs to be > >> moved outside of the locks. > > > > Ah okay, you meant the debugfs core. I see now, thanks. > > > > This indeed needs a capital solution. > > It's not obvious how to fix it.. we can probably remove taking the > list_mutex from lock_dependent(), but this still won't help the case of > memory reclaiming because reclaim may cause touching the already locked > regulator. IIUC, the case of memory reclaiming under regulator lock was > always dangerous and happened to work by chance before, correct?
I just noticed that locking in regulator_resolve_coupling() is bogus. This all holds up because regulator_list_mutex is held during the call. Feel free to test a patch below.
I'm working my way to push allocations outside of the locks, but the coupling-related locking will need to be fixed regardless.
Best Regards, Michał Mirosław
---->8<----
[PATCH] regulator: remove superfluous lock in regulator_resolve_coupling()
The code modifies rdev, but locks c_rdev instead. The bug remains: stored c_rdev could be freed just after unlock anyway. This doesn't blow up because regulator_list_mutex taken outside holds it together.
Signed-off-by: Michał Mirosław <mirq-linux@rere.qmqm.pl> --- drivers/regulator/core.c | 4 ---- 1 file changed, 4 deletions(-)
diff --git a/drivers/regulator/core.c b/drivers/regulator/core.c index 94f9225869da..e519bc9a860d 100644 --- a/drivers/regulator/core.c +++ b/drivers/regulator/core.c @@ -4859,13 +4859,9 @@ static void regulator_resolve_coupling(struct regulator_dev *rdev) return; } - regulator_lock(c_rdev); - c_desc->coupled_rdevs[i] = c_rdev; c_desc->n_resolved++; - regulator_unlock(c_rdev); - regulator_resolve_coupling(c_rdev); } } -- 2.20.1
| |