Message-ID: <CAPDyKFrN0t6Jr-gAuU4_OHEvYap2fRiaUbZC5E7sAGAQOUQjjw@mail.gmail.com>
Date:   Thu, 1 Mar 2018 12:22:14 +0100
From:   Ulf Hansson <ulf.hansson@...aro.org>
To:     Geert Uytterhoeven <geert@...ux-m68k.org>
Cc:     JeffyChen <jeffy.chen@...k-chips.com>,
        Tomasz Figa <tfiga@...omium.org>,
        Linux Kernel Mailing List <linux-kernel@...r.kernel.org>,
        Dmitry Torokhov <dmitry.torokhov@...il.com>,
        Robin Murphy <robin.murphy@....com>,
        Heiko Stuebner <heiko@...ech.de>,
        Caesar Wang <wxt@...k-chips.com>,
        Elaine Zhang <zhangqing@...k-chips.com>,
        "open list:ARM/Rockchip SoC..." <linux-rockchip@...ts.infradead.org>,
        Geert Uytterhoeven <geert+renesas@...der.be>,
        Linux ARM <linux-arm-kernel@...ts.infradead.org>,
        Linux PM <linux-pm@...r.kernel.org>
Subject: Re: [PATCH] soc: rockchip: power-domain: remove PM clocks

On 1 March 2018 at 11:37, Geert Uytterhoeven <geert@...ux-m68k.org> wrote:
> Hi Ulf,
>
> On Thu, Mar 1, 2018 at 11:18 AM, Ulf Hansson <ulf.hansson@...aro.org> wrote:
>> On 1 March 2018 at 09:33, Geert Uytterhoeven <geert@...ux-m68k.org> wrote:
>>> On Thu, Mar 1, 2018 at 4:40 AM, JeffyChen <jeffy.chen@...k-chips.com> wrote:
>>>> 3/ make sure PM clk always be controlled(otherwise it might be unexpected
>>>> disabled by other clocks under the same clk parent?):
>>>>  a) make sure Runtime PM is always enabled. and as discussed, we can select
>>>> PM in ARCH_ROCKCHIP
>>
>> I am fine enabling PM for ARCH_ROCKCHIP, if needed.
>>
>> However, what I don't agree with in general, is to make a generic
>> driver to rely on having CONFIG_PM to be set to be functional. That's
>> to me, bad practice.
>>
>> I understand, this approach exists in drivers today. I assume it
>> works, because those drivers are being used on SoCs which always has
>> CONFIG_PM set.
>
> I agree. This is mostly useful for drivers that are used on multiple SoCs,
> where the driver doesn't care about the clock (doesn't care about its
> properties), and where the clock is optional (i.e. either tied to an
> always-running bus clock, or to a controllable module clock, depending on SoC).
> The latter needs CONFIG_PM=y, but that's a platform property, controlled
> by selecting support for that specific SoC.
>
>>> On Renesas SoCs, we only add the device's module clock with pm_clk_add().
>>> Drivers that don't care about properties of the module clock just call
>>> pm_runtime_*(). That way the same driver works on different SoCs using
>>> the same device, with and without power and/or clock domains.
>>
>> I understand your point and I accept your view.
>>
>> However, I think this is more a mindset of which way one want
>> implement things. This has been discussed several times in the mailing
>> list as well.
>>
>> Surely we can cope with SoC specific constraints in drivers as well,
>> we already do that.
>>
>> In principle I think this boils done to comparing a centralized
>> method, where the PM domain deals with clocks vs a decentralized
>> method, where the driver deals with clocks. Both works, both have
>> positive and negative consequences.
>
> Is the clock (are the clocks) used purely for power management of the device,
> or does it/do they serve another (independent) purpose?

Well, I am not talking about one specific case.

I guess what you are saying is that, devices may need different kind
of clocks. Some can be power managed, some can't. That seems
reasonable.

In either way, the driver should be able to cope with both kinds of
clocks. Right!?

>
> Note that unlike clocks, power areas cannot be controlled explicitly by the
> driver. That always has to be done through the Runtime PM API.

Yes, agree!

At least until/if we get multiple power areas (PM domain) support for
devices. Then this may change. However, that's a different story. :-)

>
>> In my experience for ARM SoCs, I have found that centralized method
>> doesn't work well, when one need flexibility. For example, if there
>> are strict constraints on the order of how to put device's PM
>> resources (clocks, pinctrl, etc) in low power states. For example, to
>> avoid clock glitches.
>
> With multiple clocks used by a device, there's sometimes also a lack of
> understanding of the real clock hierarchy. If these multiple clocks need
> to be enabled/disabled in a specific order, perhaps they are not
> independent, and modelling the hierarchy correctly, and describing in DT
> the device as being connected to the leaf clock may solve the ordering issue.

Nope, this scenario I had in mind isn't about the clock hierarchy.
Instead this is about other resources that the driver deals with
during runtime PM. Pinctrl, regulators, device internals registers,
etc.

I have stumbled of cases, where the order dealing with these
resources, simply need to be strict, thus it is better managed from
the driver's runtime PM callbacks.

What I am saying is that, if you have these kind of constraints, then
having clocks being managed at the PM domain level via PM clk and
other resources in the driver, simply isn't a good mix.

>
>> Another problem with the PM clk is, more exactly with
>> pm_clk_suspend|resume(), that those invokes only clk_enable|disable().
>> pm_clk_suspend|resume() can't call clk_prepare|unprepare(), because we
>> don't know if we running in atomic context when those are executed.
>> Potentially this means leaving the clocks ungated - all the time.
>>
>> I have though about how to fix the above, several times, but I always
>> ends up with thinking that's it more easy, to let the driver deal with
>> the clocks, as then the problem goes away.
>
> There's a similar issue with powering on/off power areas.

I don't follow. Can you elaborate?

> How do device start/stop differ from power domain on/off?

I guess what you refer to is that, genpd's ->start|stop() callbacks
may be assigned to pm_clk_suspend|resume(), and those may be called
from genps's runtime PM callbacks and genpd's system sleep callbacks.

Yes, both cases suffers from the same problem as I describe above.
Clocks may stay ungated - all the time.

>
>>> Drivers that care about properties of the module clock (mainly frequency)
>>> can still use the clk_*() API for that. Other (optional) clocks must be
>>> handled by the device driver itself.
>>
>> A comment on that;
>>
>> Before we the PM clk was introduced, we didn't have the clk_bulk_*()
>> interface. To me, using clk_bulk_*() in drivers could help to simplify
>> the code in regards to manage clocks (including SoC specific clocks)
>> during runtime PM.
>>
>> Perhaps this could be an option to using PM clk, as it provides both
>> flexibility and could manage SoC specific clocks.
>
> See my comment about multiple clocks above...
>

Kind regards
Uffe

Powered by blists - more mailing lists