Re: [PATCHv2 8/8] devfreq: exynos4: Add busfreq driver for exynos4210/exynos4x12

From: Tomasz Figa
Date: Wed Mar 19 2014 - 06:23:47 EST


Hi Chanwoo,

On 19.03.2014 10:47, Chanwoo Choi wrote:
Hi Tomasz,

On 03/19/2014 12:46 AM, Tomasz Figa wrote:
On 17.03.2014 06:19, Chanwoo Choi wrote:
Hi Tomasz,

On 03/15/2014 02:35 AM, Tomasz Figa wrote:
Hi Chanwoo, Mark,

On 14.03.2014 11:56, Chanwoo Choi wrote:
Hi Mark,

On 03/14/2014 07:35 PM, Mark Rutland wrote:
On Fri, Mar 14, 2014 at 07:14:37AM +0000, Chanwoo Choi wrote:
Hi Mark,

On 03/14/2014 02:53 AM, Mark Rutland wrote:
On Thu, Mar 13, 2014 at 08:17:29AM +0000, Chanwoo Choi wrote:
This patch add busfreq driver for Exynos4210/Exynos4x12 memory interface
and bus to support DVFS(Dynamic Voltage Frequency Scaling) according to PPMU
counters. PPMU (Performance Profiling Monitorings Units) of Exynos4 SoC provides
PPMU counters for DMC(Dynamic Memory Controller) to check memory bus utilization
and then busfreq driver adjusts dynamically the operating frequency/voltage
by using DEVFREQ Subsystem.

Signed-off-by: Chanwoo Choi <cw00.choi@xxxxxxxxxxx>
---
.../devicetree/bindings/devfreq/exynos4_bus.txt | 49 ++++++++++++++++++++++
1 file changed, 49 insertions(+)
create mode 100644 Documentation/devicetree/bindings/devfreq/exynos4_bus.txt

diff --git a/Documentation/devicetree/bindings/devfreq/exynos4_bus.txt b/Documentation/devicetree/bindings/devfreq/exynos4_bus.txt
new file mode 100644
index 0000000..2a83fcc
--- /dev/null
+++ b/Documentation/devicetree/bindings/devfreq/exynos4_bus.txt
@@ -0,0 +1,49 @@
+
+Exynos4210/4x12 busfreq driver
+-----------------------------
+
+Exynos4210/4x12 Soc busfreq driver with devfreq for Memory bus frequency/voltage
+scaling according to PPMU counters of memory controllers
+
+Required properties:
+- compatible : should contain Exynos4 SoC type as follwoing:
+ - "samsung,exynos4x12-busfreq" for Exynos4x12
+ - "samsung,exynos4210-busfreq" for Exynos4210

Is there a device called "busfreq"? What device does this binding
describe?

I'll add detailed description of busfreq as following:

"busfreq(bus frequendcy)" driver means that busfreq driver control dynamically
memory bus frequency/voltage by checking memory bus utilization to optimize
power-consumption. When checking memeory bus utilization, exynos4_busfreq driver
would use PPMU(Performance Profiling Monitoring Units).

This still sounds like a description of the _driver_, not the _device_.
The binding should describe the hardware, now the high level abstraction
that software is going to build atop of it.

It sounds like this is a binding for the DMC PPMU?

Is the PPMU a component of the DMC, or is it bolted on the side?

PPMU(Performance Profiling Monitoring Unit) is to profile performance event of
various IP on Exynos4. Each PPMU provide perforamnce event for each IP.
We can check various PPMU as following:

PPMU_3D
PPMU_ACP
PPMU_CAMIF
PPMU_CPU
PPMU_DMC0
PPMU_DMC1
PPMU_FSYS
PPMU_IMAGE
PPMU_LCD0
PPMU_LCD1
PPMU_MFC_L
PPMU_MFC_R
PPMU_TV
PPMU_LEFT_BUS
PPMU_RIGHT_BUS

DMC (Dynamic Memory Controller) control the operation of DRAM in Exynos4 SoC.
If we need to get memory bust utilization of DMC, we can get memory bus utilization
from PPMU_DMC0/PPMU_DMC1.

So, Exynos4's busfreq used two(PPMU_DMC0/PPMU_DMC1) among upper various PPMU list.

Well, PPMUs and DMCs are separate hardware blocks found inside Exynos SoCs. Busfreq/devfreq is just a Linux-specific abstraction responsible for collecting data using PPMUs and controlling frequencies and voltages of appropriate power planes, vdd_int responsible for powering DMC0 and DMC1 blocks in this case.


I knew already.

I'm afraid that the binding you're proposing is unfortunately incorrect, because it represents the software abstraction, not the real hardware.

What is exactly incorrect part of this patch?


Device tree contains information about hardware, not about OS-specific drivers or subsystems. Busfreq/devfreq is not a hardware block, but a Linux-specific driver, so it's not suitable to be described by DT directly, especially considering the fact that in future it might be replaced by or merged with another subsystem.

Only PPMUs are real hardware blocks present in the SoC. In addition, a SoC-level aspect of hardware description may be added, such as a list of power planes. See below for explanation.


You means that PPMU must need separate framework from devfreq subsystem?
If PPMU framework will be implemented, PPMU framework provides API to external device driver
as common clk framework, regulator framework.
And then,exynos4 busfreq with devfreq have to use PPMU framework to monitor memory utilization.

Is this right?


Not exactly. I'm just saying that in device tree, particular hardware blocks (such as PPMU) must be represented by their own nodes. This doesn't imply any particular implementation in code. DT bindings should be considered without any particular driver code in mind.

The principle here is that with device tree describing all the information needed for Exynos bus frequency and power plane voltage scaling, you might later rewrite completely the code, without changing the DT.

Of course this also gives some hints about how a good implementation can look, which would be as you wrote a separate framework for PPMUs. However calling this a framework would be a bit exaggerated, because it can be as simple as:

- PPMU being separate platform devices, tracked by a minimal framework (e.g. a linked list and add(), remove(), find_by_node() functions),

- devfreq for given power plane also being a platform device, binding directly to power plane node, i.e. matching the compatible string of Exynos power plane,

- devfreq reading the list of PPMUs by phandles provided by its power plane node and looking up in the linked list mentioned above by find_by_node() function. Here deferred probe should happen if requested PPMU is not probed yet.

Basically that's all, the rest of the driver (except other issues mentioned in review) could stay as is.


Instead, this should be separated into several independent bindings:

- PPMU bindings to list all the PPMU instances present in the SoC and resources they need,

- power plane bindings, which define a power plane in which multiple IP blocks might reside, can be monitored by one or more PPMU units and frequency and voltage of which can be configured according to determined performance level. Needed resources will be clocks and regulators to scale and probably also operating points.

Then, exynos-busfreq driver should bind to such power planes, parse necessary data from DT (list of PPMUs and IP blocks, clocks, regulators and operating points) and register a devfreq entity.

What is 'power plane'? I don't know 'power plane'.

Power plane is a part of the SoC (set of IP blocks and buses) that is powered by the same power source. For example in Exynos4412 a power plane would be VDD_ARM, containing 4 Cortex A9 cores, two L2 cache blocks, L2 cache controller, SCU, and several other blocks. Another example in Exynos4412 would be VDD_INT plane, which contains MFC, AudioSS, LCD0, ISP, TV, CAM and several other blocks.

In addition, such power plane might contain certain facilities, such as PPMU blocks, that let you estimate bus bandwidth utilization and scale certain performance and power parameters of such power plane, e.g. frequencies of internal buses and voltage of external regulator used to power it.

If you suggest 'power plane' concept and then merge this concept to mainline,
After merged 'power plane' concept, I will apply 'power plane' concept to Exynos4's busfreq driver.

I prefer to handle 'power plane' concept on other patchset for Exynos4's busfreq driver.

This is a necessary prerequisite for having Exynos busfreq facilities described properly in device tree. This is not a concept of high level generic framework, but just a concept of hardware description that can be parsed by Linux drivers, such as Exynos busfreq.

As I said above, a requirement is that DT does not describe Linux-specific internals, but the hardware that is used by them.


exynos4 busfreq with devfreq must need to execute .probe().
So, exynos4 busfreq with devfreq should add dt node for exynos4 busfreq.
But,
You means that exynos4 busfreq can't add dt node of exynos4 busfreq
because this driver has not the hardware block. ?

Yes, this is what I mean. DT should not contain virtual nor implementation-specific data, but rather mostly generic hardware description that can be reused by any implementation of given functionality.

However in this case, what is currently a single busfreq instance would correspond to a single power plane, which is a physical hardware entity and to which the exynos4 busfreq driver can bind.

Best regards,
Tomasz
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/