Re: [PATCH V3 1/9] PM / OPP: Reword binding supporting multiple regulators per device

From: Dave Gerlach
Date: Tue Nov 15 2016 - 17:11:46 EST


Hi,
On 11/15/2016 12:56 PM, Stephen Boyd wrote:
On 11/15, Viresh Kumar wrote:
On 14-11-16, 18:13, Stephen Boyd wrote:
On 11/14, Rob Herring wrote:
On Fri, Nov 11, 2016 at 08:41:20AM +0530, Viresh Kumar wrote:
On 10-11-16, 14:51, Stephen Boyd wrote:

No. The supply names (and also clock names/index) should be left
up to the consumer of the OPP table. We don't want to encode any
sort of details like this between the OPP table and the consumer
of it in DT because then it seriously couples the OPP table to
the consumer device. "The binding" in this case that needs to be
updated is the consumer binding, to indicate that it correlated
foo-supply and bar-supply to index 0 and 1 of the OPP table
voltages.

Are you saying that we shall have a property like this then?

diff --git a/Documentation/devicetree/bindings/opp/opp.txt b/Documentation/devicetree/bindings/opp/opp.txt
index ee91cbdd95ee..733946df2fb8 100644
--- a/Documentation/devicetree/bindings/opp/opp.txt
+++ b/Documentation/devicetree/bindings/opp/opp.txt
@@ -389,7 +389,10 @@ Example 4: Handling multiple regulators
compatible = "arm,cortex-a7";
...

- cpu-supply = <&cpu_supply0>, <&cpu_supply1>, <&cpu_supply2>;
+ vcc0-supply = <&cpu_supply0>;
+ vcc1-supply = <&cpu_supply1>;
+ vcc2-supply = <&cpu_supply2>;
+ opp-supply-names = "vcc0", "vcc1", "vcc2";

Uh, no. You already have the names in the *-supply properties. Yes, they
are a PIA to retrieve compared to a *-names property, but that is the
nature of this style of binding.

Its not just PIA, but impossible AFAICT.

There are two important pieces of information we need for multiple
regulator support:
- Which regulator in the consumer node corresponds to which entry in
the OPP table. As Mark mentioned earlier, DT should be able to get
us this.

This is also possible from C code though. Or is there some case
where it isn't possible if we're sharing the same table with two
devices? I'm lost on when this would ever happen.

It feels like trying to keep the OPP table agnostic of the
consuming device and the device's binding is more trouble than
it's worth. Especially considering we have opp-shared and *-name
now.

I agree with this, I do not like having to pass a list of regulator names to the opp core that I *hope* the device I am controlling has provided. The intent seems to be to use the cpufreq-dt driver as is and not pass any cpu-supply anymore so the cpufreq-dt driver has no knowledge of what regulators are present (it operates as it would today on a system with no regulator required). But as is it will move forward regardless of whether or not we actually intended to provide a multi regulator set up or platform set_opp helper, and this probably isn't ideal. I would think cpufreq-dt/opp core should be have knowledge of what regulators are needed to achieve these opp transitions and make sure everything is in place before moving ahead.


- The order in which the supplies need to be programmed. We have all
agreed to do this in code instead of inferring it from DT and this
patch series already does that.

Agreed. Encoding a sequence into DT doesn't sound very feasible.
How is this going to be handled though? I don't see any users of
the code we're reviewing here, so it's hard to grasp how things
will work. It would be really useful if we had some user of the
code included in the patch series to get the big picture.

I have sent a patch in reply to the cover letter of this series showing the driver that I used to test multi regulator on TI am57x platform and wrote as much detail as I could on how I used what Viresh has provided. Perhaps that will show how this can be used and help to see what's missing from the core implementation here.

Previous discussions drove me to pass regulators and necessary values in the DT but do all sequencing from the driver from fixed code without inferring anything from the device tree.

Regards,
Dave



I want to solve the first problem here and I don't see how it can be
solved using such entries:

cpus {
cpu@0 {
compatible = "arm,cortex-a7";
...

vcc0-supply = <&cpu_supply0>;
vcc1-supply = <&cpu_supply1>;
vcc2-supply = <&cpu_supply2>;
operating-points-v2 = <&cpu0_opp_table>;
};
};

cpu0_opp_table: opp_table0 {
compatible = "operating-points-v2";
opp-shared;

opp@1000000000 {
opp-hz = /bits/ 64 <1000000000>;
opp-microvolt = <970000>, /* Supply 0 */
<960000>, /* Supply 1 */
<960000>; /* Supply 2 */
};
};

The code can't figure out which of vcc0, vcc1, vcc2 is added first in
the CPU node and so we need to get the order somehow. A separate
binding as I mentioned earlier is a probably (ugly) solution.

I think the problem is that Viresh wants the binding to be "self
describing" so that the OPP can be used without a driver knowing
that a supply corresponds to a particular column in the voltage
table.

Right, and that's what Mark suggested as well.

I don't understand that though. Can't we set the supply
names from C code somewhere based on the consumer of the OPPs?

That's what this patch series is doing right now.

So, are you saying that the way this patchset does it is fine with you
?

That's just to handle the ordering of operations? I need to take
a minute and understand what's changing. You may have spent
plenty of time developing/updating, but I haven't spent near
enough time understanding what's going on in these patches to
give a thorough review.