Message-ID: <CAL_Jsq+9oEJJtHC4Gp_mWy_wJMLXsv2jgqKCMK8jm1MSWT-LRg@mail.gmail.com>
Date:   Wed, 25 Jan 2017 16:32:54 -0600
From:   Rob Herring <robh@...nel.org>
To:     Dave Gerlach <d-gerlach@...com>
Cc:     Ulf Hansson <ulf.hansson@...aro.org>,
        Kevin Hilman <khilman@...libre.com>,
        Tero Kristo <t-kristo@...com>,
        "Rafael J . Wysocki" <rjw@...ysocki.net>,
        "linux-arm-kernel@...ts.infradead.org" 
        <linux-arm-kernel@...ts.infradead.org>,
        "linux-kernel@...r.kernel.org" <linux-kernel@...r.kernel.org>,
        "linux-pm@...r.kernel.org" <linux-pm@...r.kernel.org>,
        "devicetree@...r.kernel.org" <devicetree@...r.kernel.org>,
        Nishanth Menon <nm@...com>, Keerthy <j-keerthy@...com>,
        Russell King <rmk+kernel@...linux.org.uk>,
        Sudeep Holla <sudeep.holla@....com>,
        Santosh Shilimkar <ssantosh@...nel.org>,
        Lokesh Vutla <lokeshvutla@...com>
Subject: Re: [PATCH v3 2/4] dt-bindings: Add TI SCI PM Domains

On Wed, Jan 25, 2017 at 10:59 AM, Dave Gerlach <d-gerlach@...com> wrote:
> On 01/24/2017 04:03 AM, Ulf Hansson wrote:
>>
>> On 23 January 2017 at 21:11, Dave Gerlach <d-gerlach@...com> wrote:
>>>
>>> On 01/20/2017 10:52 AM, Ulf Hansson wrote:
>>>>
>>>>
>>>> [...]
>>>>
>>>>>>> Another option is create something new either common or TI SCI
>>>>>>> specific. It could be just a table of ids and phandles in the SCI
>>>>>>> node. I'm much more comfortable with an isolated property in one node
>>>>>>> than something scattered throughout the DT.
>>>>>>
>>>>>>
>>>>>>
>>>>>> To me, this seems like the best possible solution.
>>>>>>
>>>>>> However, perhaps we should also consider the SCPI Generic power domain
>>>>>> (drivers/firmware/scpi_pm_domain.c), because I believe it's closely
>>>>>> related.
>>>>>> To change the power state of a device, this PM domain calls
>>>>>> scpi_device_set|get_power_state() (drivers/firmware/arm_scpi.c), which
>>>>>> also needs a device id as a parameter. Very similar to our case with
>>>>>> the TI SCI domain.
>>>>>>
>>>>>> Currently these SCPI device ids lacks corresponding DT bindings, so
>>>>>> the scpi_pm_domain tries to work around it by assigning ids
>>>>>> dynamically at genpd creation time.
>>>>>>
>>>>>> That makes me wonder, whether we should think of something
>>>>>> common/generic?
>>>>>
>>>>>
>>>>>
>>>>> When you say something common/generic, do you mean a better binding for
>>>>> genpd,
>>>>> or something bigger than that like a new driver? Because I do think a
>>>>> phandle
>>>>> cell left open for the genpd provider to interpret solves both the scpi
>>>>> and
>>>>> ti-sci problem we are facing here in the best way. Using generic PM
>>>>> domains lets
>>>>> us do exactly what we want apart from interpreting the phandle cell
>>>>> with
>>>>> our
>>>>> driver, and I feel like anything else we try at this point is just
>>>>> going
>>>>> to be
>>>>> to work around that. Is bringing back genpd xlate something we can
>>>>> discuss?
>>>>
>>>>
>>>>
>>>> Bringing back xlate, how would that help? Wouldn't that just mean that
>>>> you will get one genpd per device? That's not an option, I think we
>>>> are all in agreement to that.
>>>
>>>
>>>
>>> Sure, perhaps the custom xlate wouldn't be the right way to do it, as we
>>> wouldn't be able to associate a device directly to a phandle, at least
>>> with
>>> how it was implemented before, but I think we can skip that entirely.
>>> Does
>>> opening up the interpretation of the cells of the 'power-domains' phandle
>>> not solve all of these issues? Is that out of the question?
>>>
>>> genpd_xlate_simple currently just makes sure the args_count of the
>>> 'power-domains' phandle was zero and bails if it was not. Why couldn't we
>>> remove this check and let the driver interpret it while still using
>>> of_genpd_add_provider_simple to register the provider? It's still a
>>> 'simple'
>>> provider from the perspective of the genpd framework and the actual pm
>>> domain mapping will not change, but now the driver can parse the cells
>>> and
>>> do whatever it needs to, such as reading a device id.
>>>
>>> I think that's a bit more flexible and will avoid breaking anything that
>>> is
>>> there today.
>>
>>
>> Would you mind providing an example? Perhaps also some code snippets
>> dealing with the parsing?
>
>
> So again the goal of this is to move the ti,sci-id value back to
> power-domains phandle instead of having a separate property, so that would
> be step one in the DT. Then in the power-domains node change
> #power-domain-cells to one. And then from there, the only change to the
> genpd framework is this:

I'd still like to understand how the ID is used in order to understand
if as a power-domain cell is appropriate. I think the test is this: is
the ID meaningful to (or defined by) the device or the power domain
controller? The former should be a property in the device node. The
latter should be phandle args.

> diff --git a/drivers/base/power/domain.c b/drivers/base/power/domain.c
> index a5e1262b964b..b82e61f0bcfa 100644
> --- a/drivers/base/power/domain.c
> +++ b/drivers/base/power/domain.c
> @@ -1603,8 +1603,6 @@ static struct generic_pm_domain genpd_xlate_simple
>                                       struct of_phandle_args *genpdspec,
>                                       void *data)
>  {
> -       if (genpdspec->args_count != 0)
> -               return ERR_PTR(-EINVAL);
>         return data;
>  }
>
>
> because genpd_xlate_simple only checks that the phandle is zero so that it
> can fail if it is not, but there's no functional reason it needs to do this.
> The genpd framework works as it did before no matter what the cells are set
> to if using of_genpd_add_provider_simple. Then in the attach_dev callback
> inside the ti_sci_pm_domains driver instead of doing
>
>         ret = of_property_read_u32(np, "ti,sci-id", &idx);
>
> to read the ti,sc-id for a device into idx we can now do:
>
>        ret = of_parse_phandle_with_args(np, "power-domains",
>                                    "#power-domain-cells", 0, &pd_args);
>        idx = pd_args.args[0];
>
> or even simpler from within our driver
>
>         ret = of_property_read_u32_index(np, "power-domains", 1, &idx);

This you should not be doing. The client driver shouldn't care how
many cells or what their values are.

Rob

Powered by blists - more mailing lists