Re: [EXT] Re: [PATCH v3] mmc: sdhci-xenon: Add Xenon SDHCI specific system-level PM support

[Zhoujie Wu <zjwu@marvell.com>]

On 07/14/2017 02:09 AM, Ulf Hansson wrote:
> On 13 July 2017 at 23:45, Zhoujie Wu <zjwu@marvell.com> wrote:
>>
>> On 07/13/2017 04:03 AM, Ulf Hansson wrote:
>>> On 13 July 2017 at 12:48, Ulf Hansson <ulf.hansson@linaro.org> wrote:
>>>> On 13 July 2017 at 12:13, Jisheng Zhang <jszhang@marvell.com> wrote:
>>>>> On Thu, 13 Jul 2017 11:52:54 +0200 Ulf Hansson wrote:
>>>>>
>>>>>> On 13 July 2017 at 11:25, Jisheng Zhang <jszhang@marvell.com> wrote:
>>>>>>> Hi Ulf,
>>>>>>>
>>>>>>> On Thu, 13 Jul 2017 11:18:32 +0200 Ulf Hansson wrote:
>>>>>>>
>>>>>>>> On 13 July 2017 at 00:16, Zhoujie Wu <zjwu@marvell.com> wrote:
>>>>>>>>> From: Hu Ziji <huziji@marvell.com>
>>>>>>>>>
>>>>>>>>> Add Xenon specific system-level suspend and resume support.
>>>>>>>>> Especially during resume, re-configure Xenon specific registers
>>>>>>>>> since registers setting will be lost in suspend if Xenon is power
>>>>>>>>> off.
>>>>>>>> I recommend to start with deploying runtime PM support instead of
>>>>>>>> system PM support. Then on top of such change, you should make use of
>>>>>>>> the runtime PM centric path to get system sleep support for "free"
>>>>>>>> (and thus all the nice benefits).
>>>>>>> I'm not sure whether runtime PM is useful for xenon case. The xenon HW
>>>>>>> support ACG(Auto Clock Gating) and SDCLK-Off-While-Idle features,
>>>>>>> that's
>>>>>>> to say we even don't need to do anything but achieve the runtime PM
>>>>>>> gains.
>>>>>> Yeah, but that's only internally managed by mmc controller. The clock
>>>>>> will not be unprepared/disabled, from clock tree point of view. Isn't
>>>>>> that also worth doing?
>>>>>>
>>>>> The HW is clock gated, the difference is clock itself. From power saving
>>>>> point of view, the gain is nearly zero. From latency point of view,
>>>>> could
>>>> I assume the clock you are talking about is the "core" clock? I then
>>>> assumes that clock is used as the interface clock for the card?
>>>>
>>>> That makes me wonder, don't you have other device clocks to manage as
>>>> well? Clocks that is provided to the controller to make it functional?
>> At first, really appreciate your quick and valuable feedback.
>> The core clock in this driver is the clock provided by SOC to sdh
>> controller, and there is a divider inside the controller to generate sdclk
>> which provides to sd/emmc card.
>> Actually there are two runtime power saving features inside the controller
>> per my understanding.
>> sdclk_idle_enable will cut the clock to sd/emmc card if sd bus idle for some
>> time. auto_clkgate_enable means HW will auto gate the clock to sdh
>> controller core logic.
> I am not sure I get the second part here. The clock to shd is enabled
> via a call to clk_prepare_enable(). Unless you explicitly call
> clk_disable_unprepare() for it, no? How can any outer logic know when
> it can be gated?

This is my understanding. Hope it can make you clear.
The clock tree is like below.
SOC --> [ SDH_CLK_GEN --> SDH_CONTROLLER ] --> SD/EMMC CARD

There is one clock generator inside sdh slot IP, SOC provides the clock 
to the sdh slot IP. This clock is enabled/disabled by SW when calling 
clk_prepare_enable/clk_disable_unprepare.
The auto clock gating is not any outer logic, it is inside sdh slot IP, 
when sdh controller has no activity, the IP will gate the clock from 
sdh_clk_gen to sdh_controller.  sdh_clk_gen logic itself still has clock 
from SOC.
With or without runtime pm, the only difference is if sdh_clk_gen has 
clock or not. So the power benefit is limited.

>
>> With SW runtime pm mechanism, compares with HW auto clock gating, the only
>> difference is SW cut the source of sdh clock tree, external clock gating vs
>> internal clock gating, there will be some benefits, but limited.
> Right.
>
>> Previously we enabled the runtime pm mechanism in our mobile products, which
>> were using the same IP(some old version, including 3 sdh slots) with auto
>> clock gating feature(the driver is sdhci-pxav3.c).  The saving of power was
>> about 2~3mA@vcc_main_1.05V(28nm chip) with 3 sdh slots inside soc. No more
>> than 1mA/1sdh slot.
> 1 mA/sdh slot is a great reason to deploy runtime PM support. For a
> battery driven device that would be a significant improvement.
>
> Back in the days when I worked at ST-Ericssion, we were chasing uA
> when optimizing for power-save. :-)

Definitely for mobile products, but now I didn't see urgent requirement 
for our networking products.
>
>> I read sdhci-of-at91 driver and your recommended patch, I got your point is
>> using a light way for system sleep based on runtime pm feature. From SW
>> perspective, kill two birds with one stone, it is good.
> Right.
>
>> But considering about the benefits, it is not that urgent to take runtime pm
>> feature as a must, it is a better to have feature. System standby is a must
>> feature, without this patch, the system can't work well after resume.
>> Do you think it is reasonable to add complete standby support at first, then
>> take runtime pm as a next step?
> You can do that, but why? And will then the "next step" ever happen?
>
> Do you really want to spend efforts in getting something working for
> system suspend only, while you instead easily could deploy both
> runtime PM and system PM support at the same time?

As Ziji said in another mail, it takes time for next step. The runtime 
pm need to be verified completely on all supported boards.
I understand from SW perspective, we'd better have both. But I need 
input from internal customers to see if they only request system sleep 
or they want both, and what's their priority.
>
>>> Besides the clocks, you have the xenon mmc phy. Can't that also be put
>>> that in some low power mode at request in-activity?
>> For the phy behavior, currently I don't see any SW operation for the lpm, I
>> will check with HW guys about the behaviour.
> Great, that would really be interesting from a runtime PM point of view.
>
> Perhaps then also ask if re-configuring the phy via xenon_phy_adj(),
> makes sense when powering off the card? Because currently you seems to
> keep the latest configuration, even if the mmc core decides to power
> off the card during system sleep. Unless I am reading the code wrong
> from the ->set_ios() ops.
> Kind regards
> Uffe