Received-SPF: pass (google.com: best guess record for domain of linux-kernel-owner@vger.kernel.org designates 209.132.180.67 as permitted sender) client-ip=209.132.180.67;
Message-ID: <62557522be4924a01d3822d4734c30f2965c608b.camel@partner.samsung.com>
Subject: Re: [RFC PATCH 09/11] devfreq: exynos-bus: Add interconnect
 functionality to exynos-bus
From:   Artur =?UTF-8?Q?=C5=9Awigo=C5=84?= <a.swigon@partner.samsung.com>
To:     Georgi Djakov <georgi.djakov@linaro.org>,
        devicetree@vger.kernel.org, linux-arm-kernel@lists.infradead.org,
        linux-samsung-soc@vger.kernel.org, linux-kernel@vger.kernel.org,
        linux-pm@vger.kernel.org, dri-devel@lists.freedesktop.org
Cc:     krzk@kernel.org, cw00.choi@samsung.com, myungjoo.ham@samsung.com,
        inki.dae@samsung.com, sw0312.kim@samsung.com,
        m.szyprowski@samsung.com,
        =?UTF-8?Q?Bart=C5=82omiej_?= =?UTF-8?Q?=C5=BBo=C5=82nierkiewicz?= 
        <b.zolnierkie@samsung.com>
Date:   Thu, 01 Aug 2019 09:59:42 +0200
In-Reply-To: <6e8b2081-2fb3-9ab8-37d1-8b5fe5fd8e11@linaro.org>
Mime-Version: 1.0
Content-Transfer-Encoding: 8bit
Content-Type: text/plain; charset="utf-8"
CMS-TYPE: 201P
References: <20190723122016.30279-1-a.swigon@partner.samsung.com>
        <CGME20190723122028eucas1p2eb75f35b810e71d6c590370aaff0997b@eucas1p2.samsung.com>
        <20190723122016.30279-10-a.swigon@partner.samsung.com>
        <6e8b2081-2fb3-9ab8-37d1-8b5fe5fd8e11@linaro.org>
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk

Hi Georgi,

On Fri, 2019-07-26 at 11:05 +0300, Georgi Djakov wrote:
> Hi Artur,
> 
> On 7/23/19 15:20, Artur Świgoń wrote:
> > This patch adds interconnect functionality to the exynos-bus devfreq
> > driver.
> > 
> > The SoC topology is a graph (or, more specifically, a tree) and most of its
> > edges are taken from the devfreq parent-child hierarchy (cf.
> > Documentation/devicetree/bindings/devfreq/exynos-bus.txt). The previous
> > patch adds missing edges to the DT (under the name 'parent'). Due to
> > unspecified relative probing order, -EPROBE_DEFER may be propagated to
> > guarantee that a child is probed before its parent.
> > 
> > Each bus is now an interconnect provider and an interconnect node as well
> > (cf. Documentation/interconnect/interconnect.rst), i.e. every bus registers
> > itself as a node. Node IDs are not hardcoded but rather assigned at
> > runtime, in probing order (subject to the above-mentioned exception
> > regarding relative order). This approach allows for using this driver with
> > various Exynos SoCs.
> 
> I am not familiar with the Exynos bus topology, but it seems to me that it's not
> represented correctly. An interconnect provider with just a single node (port)
> is odd. I would expect that each provider consists of multiple master and slave
> nodes. This data would be used by a framework to understand what are the links
> and how the traffic flows between the IP blocks and through which buses.

To summarize the exynos-bus topology[1] used by the devfreq driver: There are
many data buses for data transfer in Samsung Exynos SoC. Every bus has its own
clock. Buses often share power lines, in which case one of the buses on the
power line is referred to as 'parent' (or as 'devfreq' in the DT). In the
particular case of Exynos4412[1][2], the topology can be expressed as follows:

bus_dmc
-- bus_acp
-- bus_c2c

bus_leftbus
-- bus_rightbus
-- bus_display
-- bus_fsys
-- bus_peri
-- bus_mfc

Where bus_dmc and bus_leftbus probably could be referred to as masters, and the
following indented nodes as slaves. Patch 08/11 of this RFC additionally adds
the following to the DT:

bus_dmc
-- bus_leftbus

Which makes the topology a valid tree.

The exynos-bus concept in devfreq[3] is designed in such a way that every bus is
probed separately as a platform device, and is a largely independent entity.
This RFC proposes an extension to the existing devfreq driver that basically
provides a simple QoS to ensure minimum clock frequency for selected buses
(possibly overriding devfreq governor calculations) using the interconnect
framework.

The hierarchy is modelled in such a way that every bus is an interconnect node.
On the other hand, what is considered an interconnect provider here is quite
arbitrary, but for the reasons mentioned in the above paragraph, this RFC
assumes that every bus is a provider of itself as a node. Using an alternative
singleton provider approach was deemed more complicated since the 'dev' field in
'struct icc_provider' has to be set to something meaningful and we are tied to
the 'samsung,exynos-bus' compatible string in the driver (and multiple instances
of exynos-bus probed in indeterminate relative order).

I'm looking forward to hearing any additional thoughts you may have on this
topic.

Best regards,
-- 
Artur Świgoń
Samsung R&D Institute Poland
Samsung Electronics

[1] Documentation/devicetree/bindings/devfreq/exynos-bus.txt
[2]
arch/arm/boot/dts/exynos4412-odroid-common.dtsi
[3] drivers/devfreq/exynos-bus.c
(subject of this patch)