Re: [PATCH v7 0/6] Add coupled regulators mechanism

[Dmitry Osipenko]

On 4/23/18 5:33 PM, Maciej Purski wrote:
> Hi all,
> 
> this patchset adds a new mechanism to the framework - regulators' coupling.
> 
> On Odroid XU3/4 and other Exynos5422 based boards there is a case, that
> different devices on the board are supplied by different regulators
> with non-fixed voltages. If one of these devices temporarily requires
> higher voltage, there might occur a situation that the spread between
> devices' voltages is so high, that there is a risk of changing
> 'high' and 'low' states on the interconnection between devices powered
> by those regulators.
> 
> Algorithmicaly the problem was solved by:
> Inderpal Singh <inderpal.s@xxxxxxxxxxx>
> Doug Anderson <dianders@xxxxxxxxxxxx>
> 
> The discussion on that subject can be found here:
> https://lkml.org/lkml/2014/4/29/28
> 
> Therefore this patchset is an attempt to apply the idea to regulators core
> as concluded in the discussion by Mark Brown and Doug Anderson.
> 
> This feature is required to enable support for generic CPUfreq
> and devfreq drivers for the mentioned boards.
> 
> The current assumption is that an uncoupled regulator is a special case
> of a coupled one, so they should share a common voltage setting path.


Hello Maciej,

I'm working on the CPUFreq driver for NVIDIA Tegra20/30 and now looking at adding support for DVFS to the driver. The "coupled regulators mechanism" is just what is needed for Tegra.

I'm now using and relying on the "coupled regulators" patches, here are my findings and thoughts on the patchset:

1.

There is a technical problem with resolving of coupled regulators. I'm running into a NULL dereference on a regulator_set_voltage() because one of two coupled regulators doesn't resolve the couple. Solution is to return EPROBE_DEFER on regulator requesting until the coupled pair is fully resolved and re-try the resolving for each of regulators after the regulators registration. Here is the patch:

diff --git a/drivers/regulator/core.c b/drivers/regulator/core.c
index 282511508698..baa8ef91bc30 100644
--- a/drivers/regulator/core.c
+++ b/drivers/regulator/core.c
@@ -1747,6 +1747,12 @@ struct regulator *_regulator_get(struct device *dev, const char *id,
                return regulator;
        }
 
+       if (rdev->coupling_desc.n_resolved != rdev->coupling_desc.n_coupled) {
+               regulator = ERR_PTR(-EPROBE_DEFER);
+               put_device(&rdev->dev);
+               return regulator;
+       }
+
        ret = regulator_resolve_supply(rdev);
        if (ret < 0) {
                regulator = ERR_PTR(ret);
@@ -4433,6 +4439,9 @@ static int regulator_register_fill_coupling_array(struct device *dev,
        if (!IS_ENABLED(CONFIG_OF))
                return 0;
 
+       if (rdev == data)
+               return 0;
+
        if (regulator_fill_coupling_array(rdev))
                rdev_dbg(rdev, "unable to resolve coupling\n");
 
@@ -4663,6 +4672,11 @@ regulator_register(const struct regulator_desc *regulator_desc,
        /* try to resolve regulators supply since a new one was registered */
        class_for_each_device(&regulator_class, NULL, NULL,
                              regulator_register_resolve_supply);
+
+       /* try to resolve regulators coupling since a new one was registered */
+       class_for_each_device(&regulator_class, NULL, rdev,
+                             regulator_register_fill_coupling_array);
+
        kfree(config);
        return rdev;

2.

It looks to me that definition of the max-spread property per groups of regulators isn't future-proof and makes it uneasy to extend the coupling properties later. Much better should be to define the max-spread property per regulators pair, like in this example:

regA: regulatorA {
	regulator-coupled-with = <&regB &regC>;
	regulator-coupled-max-spread = <100000 200000;
};

regB: regulatorB {
	regulator-coupled-with = <&regA &regC>;
	regulator-coupled-max-spread = <100000 300000>;
};

regC: regulatorC {
	regulator-coupled-with = <&regA &regB>;
	regulator-coupled-max-spread = <200000 300000>;
};

Right now we can just change the binding, requiring max-spread values to be set per-pair and not allowing spread values to differ in the code, hence only simple cases of coupling will be supported for the starter.

3.

The patches were reverted in upstream because they caused regression on some board, at least on Tegra I haven't noticed any problems with the locking. Are you planing to get back to working on the patches anytime soon?