Re: [PATCH 1/5] regulator: core: Resolve supply earlier

[Thierry Reding <thierry.reding@gmail.com>]

[-- Attachment #1: Type: text/plain, Size: 2956 bytes --]

On Mon, Apr 11, 2016 at 11:59:02AM +0100, Jon Hunter wrote:
> Hi Thierry,
> 
> On 07/04/16 15:22, Thierry Reding wrote:
> > From: Thierry Reding <treding@nvidia.com>
> > 
> > Subsequent patches will need access to the parent supply from within the
> > set_machine_constraints() function to properly implement bypass mode. If
> > the parent supply hasn't been resolved by that time the voltage can't be
> > queried.
> > 
> > Also, by making sure the supply is resolved early most of the changes in
> > set_machine_constraints() don't have to be undone if resolution fails.
> > 
> > Suggested-by: Mark Brown <broonie@kernel.org>
> > Signed-off-by: Thierry Reding <treding@nvidia.com>
> > ---
> >  drivers/regulator/core.c | 27 +++++++++++++++++++++------
> >  1 file changed, 21 insertions(+), 6 deletions(-)
> > 
> > diff --git a/drivers/regulator/core.c b/drivers/regulator/core.c
> > index 2786d251b1cc..cc0333a79924 100644
> > --- a/drivers/regulator/core.c
> > +++ b/drivers/regulator/core.c
> > @@ -3972,18 +3972,27 @@ regulator_register(const struct regulator_desc *regulator_desc,
> >  
> >  	dev_set_drvdata(&rdev->dev, rdev);
> >  
> > +	if (init_data && init_data->supply_regulator)
> > +		rdev->supply_name = init_data->supply_regulator;
> > +	else if (regulator_desc->supply_name)
> > +		rdev->supply_name = regulator_desc->supply_name;
> > +
> > +	/*
> > +	 * set_machine_constraints() needs the supply to be resolved in order
> > +	 * to support querying the current voltage in bypass mode. Resolve it
> > +	 * here to more easily handle deferred probing.
> > +	 */
> > +	ret = regulator_resolve_supply(rdev);
> > +	if (ret < 0)
> > +		goto scrub;
> > +
> 
> Thanks for sending this. However, I think that calling
> regulator_resolve_supply() can cause a deadlock, because the
> regulator_list_mutex is held at this point and
> regulator_resolve_supply() calls regulator_dev_lookup() which may try to
> request the mutex again.

True... I never encountered that case in my testing. I'm not sure
exactly why, though.

> So may be we need to move this call after the call to
> regulator_of_get_init_data() before we acquire the mutex.

I don't think that'll work. regulator_resolve_supply() depends on some
operations performed much later (such as rdev->dev.parent being set).

Perhaps moving the locking of the regulator_list_mutex down instead
could work. It seems to me like the first place where it would need to
be held is set_machine_constraints().

> Also, if we add this call, then I am wondering if we still need ...
> 
> 	class_for_each_device(&regulator_class, NULL, NULL,
> 			      regulator_register_resolve_supply);

Possibly not. That line was introduced to hook up existing orphan
regulators with their parents when they were registered, but I guess
since we now always defer probe if a parent isn't registered yet the
line would become a no-op.

Thierry

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 819 bytes --]