Re: [PATCH v2] PM / devfreq: Add support for QCOM devfreq firmware

From: skannan
Date: Fri Jul 27 2018 - 20:16:32 EST

On 2018-05-23 07:39, Rob Herring wrote:

Reviving an old thread. Sorry about the late reply. Got busy.

On Tue, May 22, 2018 at 1:30 PM, Saravana Kannan <skannan@xxxxxxxxxxxxxx> wrote:
On 05/22/2018 11:08 AM, Rob Herring wrote:

On Fri, May 18, 2018 at 12:52:40AM -0700, Saravana Kannan wrote:

The firmware present in some QCOM chipsets offloads the steps necessary
for
changing the frequency of some devices (Eg: L3). This driver implements
the
devfreq interface for this firmware so that various governors could be
used
to scale the frequency of these devices.

Each client (say cluster 0 and cluster 1) that wants to vote for a
particular device's frequency (say, L3 frequency) is represented as a
separate voter device (qcom,devfreq-fw-voter) that's a child of the
firmware device (qcom,devfreq-fw).

Signed-off-by: Saravana Kannan <skannan@xxxxxxxxxxxxxx>
---
.../bindings/devfreq/devfreq-qcom-fw.txt | 41 +++
drivers/devfreq/Kconfig | 14 +
drivers/devfreq/Makefile | 1 +
drivers/devfreq/devfreq_qcom_fw.c | 330
+++++++++++++++++++++
4 files changed, 386 insertions(+)
create mode 100644
Documentation/devicetree/bindings/devfreq/devfreq-qcom-fw.txt
create mode 100644 drivers/devfreq/devfreq_qcom_fw.c

diff --git
a/Documentation/devicetree/bindings/devfreq/devfreq-qcom-fw.txt
b/Documentation/devicetree/bindings/devfreq/devfreq-qcom-fw.txt
new file mode 100644
index 0000000..f882a0b
--- /dev/null
+++ b/Documentation/devicetree/bindings/devfreq/devfreq-qcom-fw.txt
@@ -0,0 +1,41 @@
+QCOM Devfreq firmware device
+
+Some Qualcomm Technologies, Inc. (QTI) chipsets have a firmware that
+offloads the steps for frequency switching. It provides a table of
+supported frequencies and a register to request one of the supported
+freqencies.
+
+The qcom,devfreq-fw represents this firmware as a device. Sometimes,
+multiple entities want to vote on the frequency request that is sent to
the
+firmware. The qcom,devfreq-fw-voter represents these voters as child
+devices of the corresponding qcom,devfreq-fw device.
+
+Required properties:
+- compatible: Must be "qcom,devfreq-fw" or
"qcom,devfreq-fw-voter"


No versions of firmware?


Sure, I can add a v1. Right now the interface has always been identical. I
thought if it changed in the future I'll add -v2.

Sounds like you are making up version numbers. If you don't have real
h/w or firmware version numbers, then use an SoC specific compatible
string.

+Only for qcom,devfreq-fw:
+- reg: Pairs of physical base addresses and region sizes
of
+ memory mapped registers.


Registers? Is this firmware or h/w block?


It's a HW block that has its own firmware.

So you have 2 things that could change: the h/w interface and the
firmware version. Make sure the compatible string(s) is specific
enough for the OS to know the exact combination.

For all practical purposes, the FW is opaque to the OS. It doesn't affect anything the OS can do with the IP block. So, the HW version is what matters. I'll figure out the actual HW version and use that.

+- reg-names: Names of the bases for the above registers.
+ Required register regions are:
+ - "en-base": address of register to check if the
+ firmware is enabled.
+ - "ftbl-base": address region for the frequency
+ table.
+ - "perf-base": address of register to request a
+ frequency.
+
+Example:
+
+ qcom,devfreq-l3 {
+ compatible = "qcom,devfreq-fw";
+ reg-names = "en-base", "ftbl-base", "perf-base";
+ reg = <0x18321000 0x4>, <0x18321110 0x600>, <0x18321920
0x4>;
+
+ qcom,cpu0-l3 {
+ compatible = "qcom,devfreq-fw-voter";


There's no point in these nodes. They don't have any properties or
resources.


These nodes decide how many voters this device supports. Each voter would be
a devfreq node that will have its own governor set. For example, one of them
would use this governor:
http://lkml.iu.edu/hypermail/linux/kernel/1805.2/02474.html

You can also attach different devfreq-event devices to each one of these
voter devices based on what events you want to use for scaling each voter.
So, the devices are definitely needed in the larger context.

Sorry, I still don't understand.

Ok, let me try to explain. Let's take L3 as an example. Different other IPs might have different requirements on the L3 frequency. For example, little CPUs might want L3 to run at 400 MHz, big CPUs might want L3 to run at 1000 MHz, GPU or some other peripheral might want L3 to run at 800 MHz. The L3 freq needs to be set to max of these requests -- in this case 1000 MHz. I'm trying to represent each of these "votes" on L3 as a device. Once I register these with devfreq, each of these devices will have a devfreq device created for them (devfreq framework does this) and we can have different governors for each of these voters. So, I need these child devices to represent each voter. There's no getting around needing one device per voter and having to aggregate their votes into a parent device.

Does that make sense?

Thanks,
Saravana