Re: [PATCH] drm/i2c: tda998x: Fix lockdep warning about possible circular dependency

[Daniel Vetter]

On Thu, Jul 20, 2017 at 4:40 PM, Liviu Dudau <Liviu.Dudau@xxxxxxx> wrote:
> On Thu, Jul 20, 2017 at 03:24:13PM +0100, Russell King - ARM Linux wrote:
>> On Thu, Jul 20, 2017 at 03:19:10PM +0100, Liviu Dudau wrote:
>> > On Thu, Jul 20, 2017 at 02:08:29PM +0100, Russell King - ARM Linux wrote:
>> > > On Thu, Jul 20, 2017 at 01:54:04PM +0100, Liviu Dudau wrote:
>> > > > On Thu, Jul 20, 2017 at 12:44:49PM +0100, Russell King - ARM Linux wrote:
>> > > > > Actually, scrub that idea - drm_helper_probe_single_connector_modes()
>> > > > > calls drm_edid_to_eld() for these cases anyway, so we must call
>> > > > > drm_helper_probe_single_connector_modes() with the audio_mutex held.
>> > > >
>> > > > OK, so the lockdep warning is spurious?
>> > >
>> > > I don't think so.  I think there's two ways to solve this:
>> > >
>> > > 1. replace the audio_mutex in tda998x_audio_get_eld() and
>> > >    tda998x_connector_fill_modes() with a new mutex (eld_mutex) to
>> > >    protect just the ELD.
>> > >
>> > > 2. remove the mutex from these two functions, and take the connection_mutex
>> > >    modeset lock in tda998x_audio_get_eld().
>> > >
>> > > However, I don't have a view on which would be best.
>> >
>> > If you don't mind, I took the liberty of picking option 2, just because
>> > I don't like adding new locks when existing ones might do the job.
>>
>> I don't mind - but one question for the DRM people in connection with
>> your patch is whether we need the acquire context for this relatively
>> simple lock/copy/unlock sequence.  This path for getting the ELD
>> shouldn't be holding any other DRM locks.
>
> Cc-ing Daniel Vetter in hope of clarifications / nod of approval.
> However, I can only see my emails in the online dri-devel archive, not
> yours, so I can't point him to the whole discussion.
>
> danvet: a while ago while I was debugging the delayed fb setup I found
> a lockdep warning with the tda998x driver. Now I've had some more time
> to investigate so I have created a patch trying to fix the issue, which
> was on v1 just a re-ordering of places where tda998x's audio_mutex lock
> was taken. Russell suggested a different approach, which I have
> implemented in [1], but we wonder if we really have to go through the
> whole dance.

If all you do is take only one ww mutex (wrapped up in
drm_modeset_lock for kms) then you can pass a NULL acquire context.
The context is only needed when you want to take multiple locks at the
same time (to be able to resolve deadlocks). Taking a single lock
within the modeset lock class can't deadlock.

Reading the kerneldoc that's not explained at all :-( Can you pls type
a patch to improve the docs for drm_modeset_lock?

Thanks, Daniel
-- 
Daniel Vetter
Software Engineer, Intel Corporation
+41 (0) 79 365 57 48 - http://blog.ffwll.ch