Re: [PATCH 1/3] regulator: core: fix boot-on regulators use_count usage

[Matti Vaittinen]

Hi dee Ho Peeps,

Long time no hear =)

On Tue, Oct 01, 2019 at 12:57:31PM -0700, Doug Anderson wrote:
> Hi,
> 
> On Fri, Sep 27, 2019 at 1:47 AM Marco Felsch <m.felsch@xxxxxxxxxxxxxx> wrote:
> > > > > > It should be possible to do a regulator_disable() though I'm not
> > > > > > sure anyone actually uses that.  The pattern for a regular
> > > > > > consumer should be the normal enable/disable pair to handle
> > > > > > shared usage, only an exclusive consumer should be able to use
> > > > > > just a straight disable.
> >
> > In my case it is a regulator-fixed which uses the enable/disable pair.
> > But as my descriptions says this will not work currently because boot-on
> > marked regulators can't be disabled right now (using the same logic as
> > always-on regulators).
> >

I was developing driver for yet-another ROHM PMIC when I hit the
phenomena you have been discussing here (I think) :) I used regulator-boot-on
flag from DT in my test setup and then did a test consumer who does

regulator_get()
regulator_enable()
regulator_disable() pair.

As this 'test consumer' was only user for the regulator I expected the
regulator to be disabled after call to regulator_disable. But it was
not.

It seems to me that the use_count is incremented for boot-on regulators
before first call to regulator_enable(). So when the consumer does first
regulator_enable() the use_count will actually go to 2. Hence the
corresponding regulator_disable() won't actually disable the regulator
even though the consumer is actually only known user.

I did unbalanced regulator_disable() - which does disable the regulator
but it also spills the warning.

I did instrument the regmap helpers and regulator_enable/disable to
dump out the actual i2c accesses and use_counts. Regulator enable prints
use_count _before_ incrementing it.


Check enable state after regulator_get (calls regulator_is_enabled)
root@arm:/sys/kernel/mva_test/regulators# cat buck3_en 
[  123.251499] dbg_regulator_is_enabled_regmap: called for 'buck3'
[  123.257524] regulator_is_enabled_regmap_dbg: Reading reg 0x1c
[  123.267386] regulator_is_enabled_regmap_dbg: read succeeded, val 0xe

Enable regulator by test consumer (no i2c access as regulator is on)
1root@arm:/sys/kernel/mva_test/regulators# echo 1 > buck3_en 
[  171.438524] Calling regulator_enable
[  171.446324] Enable requested, use-count 1

/* disable regulator by consumer */
root@arm:/sys/kernel/mva_test/regulators# echo 0 > buck3_en                                                                                     
[  187.799956] Calling regulator_disable
[  187.805935] regulator disable requested, use-count 2, always-on 0

/* Unbalanced disble */
root@arm:/sys/kernel/mva_test/regulators# echo 0 > buck3_en 
[  207.832682] Calling regulator_disable
[  207.842949] regulator disable requested, use-count 1, always-on 0
[  207.849237] regulator do disable
[  207.852502] dbg_regulator_disable_regmap: called for 'buck3'
[  207.858272] regulator_disable_regmap_dbg: reg 0x1c mask 0x8 val 0x0, masked_val 0x0
[  207.909942] buck3: Underflow of regulator enable count
[  207.915189] Failed to toggle regulator state. error(-22)
bash: echo: write error: Invalid argument
root@arm:/sys/kernel/mva_test/regulators# 

> > > > > Ah, I see, I wasn't aware of the "exclusive" special case!  Marco: is
> > > > > this working for you?  I wonder if we need to match
> > > > > "regulator->enable_count" to "rdev->use_count" at the end of
> > > > > _regulator_get() in the exclusive case...
> >
> > So my fix isn't correct to fix this in general?
> 
> I don't think your fix is correct.  It sounds as if the intention of
> "regulator-boot-on" is to have the OS turn the regulator on at bootup
> and it keep an implicit reference until someone explicitly tells the
> OS to drop the reference.

Hmm.. What is the intended way to explicitly tell the OS to drop the
reference? I would assume we should still use same logic as with other
regulators - if last user calls regulator_disable() we should disable
the regulator? (I may not understand all this well enough though)
 
> > > > Yes, I think that case has been missed when adding the enable
> > > > counts - I've never actually had a system myself that made any
> > > > use of this stuff.  It probably needs an audit of the users to
> > > > make sure nobody's relying on the current behaviour though I
> > > > can't think how they would.
> > >
> > > Marco: I'm going to assume you'll tackle this since I don't actually
> > > have any use cases that need this.
> >
> > My use case is a simple regulator-fixed which is turned on by the
> > bootloader or to be more precise by the pmic-rom. To map that correctly
> > I marked this regulator as boot-on. Unfortunately as I pointed out above
> > this is handeld the same way as always-on.

Here I am again just a man in the middle as I am "only a component vendor"
and lack of complete system information. But I _think_ some of the users
of BD71827 and BD71847 PMICs do use setup where regulator-boot-on is
used to enable certain BUCKs to power some graphics chip at start-up. At
later stage it should be possible to cut the power in order to do power
saving or decrease heating when graphichs are not needed. So I think it
would be nice to fix this somehow.

> It's a fixed regulator controlled by a GPIO?  Presumably the GPIO can
> be read.  That would mean it ideally shouldn't be using
> "regulator-boot-on" since this is _not_ a regulator whose software
> state can't be read.  Just remove the property.

How should we handle cases where we want OS to enable regulator at
boot-up - possibly before consumer drivers can be load?


Br,
	Matti Vaittinen
-- 
Matti Vaittinen, Linux device drivers
ROHM Semiconductors, Finland SWDC
Kiviharjunlenkki 1E
90220 OULU
FINLAND

~~~ "I don't think so," said Rene Descartes. Just then he vanished ~~~
Simon says - in Latin please.
~~~ "non cogito me" dixit Rene Descarte, deinde evanescavit ~~~
Thanks to Simon Glass for the translation =]