[Xen-devel] [PATCH v4] xen/arm: Add a clock property

[Dirk Behme]

On 14.07.2016 17:55, Stefano Stabellini wrote:
> On Thu, 14 Jul 2016, Dirk Behme wrote:
>> On 14.07.2016 12:38, Stefano Stabellini wrote:
>>> On Thu, 14 Jul 2016, Dirk Behme wrote:
>>>> On 13.07.2016 23:03, Michael Turquette wrote:
>>>>> Quoting Dirk Behme (2016-07-13 11:56:30)
>>>>>> On 13.07.2016 20:43, Stefano Stabellini wrote:
>>>>>>> On Wed, 13 Jul 2016, Dirk Behme wrote:
>>>>>>>> On 13.07.2016 00:26, Michael Turquette wrote:
>>>>>>>>> Quoting Dirk Behme (2016-07-12 00:46:45)
>>>>>>>>>> Clocks described by this property are reserved for use by Xen,
>>>>>>>>>> and
>>>>>>>>>> the OS
>>>>>>>>>> must not alter their state any way, such as disabling or
>>>>>>>>>> gating a
>>>>>>>>>> clock,
>>>>>>>>>> or modifying its rate. Ensuring this may impose constraints on
>>>>>>>>>> parent
>>>>>>>>>> clocks or other resources used by the clock tree.
>>>>>>>>>
>>>>>>>>> Note that clk_prepare_enable will not prevent the rate from
>>>>>>>>> changing
>>>>>>>>> (clk_set_rate) or a parent from changing (clk_set_parent). The
>>>>>>>>> only
>>>>>>>>> way
>>>>>>>>> to do this currently would be to set the following flags on the
>>>>>>>>> effected
>>>>>>>>> clocks:
>>>>>>>>>
>>>>>>>>>      CLK_SET_RATE_GATE
>>>>>>>>>      CLK_SET_PARENT_GATE
>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>> Regarding setting flags, I think we already talked about that. I
>>>>>>>> think
>>>>>>>> the
>>>>>>>> conclusion was that in our case its not possible to manipulate the
>>>>>>>> flags in
>>>>>>>> the OS as this isn't intended to be done in cases like ours.
>>>>>>>> Therefore
>>>>>>>> no API
>>>>>>>> is exported for this.
>>>>>>>>
>>>>>>>> I.e. if we need to set these flags, we have to do that in Xen
>>>>>>>> where we
>>>>>>>> add the
>>>>>>>> clocks to the hypervisor node in the device tree. And not in the
>>>>>>>> kernel patch
>>>>>>>> discussed here.
>>>>>>>
>>>>>>> These are internal Linux flags, aren't they?
>>>>>>
>>>>>>
>>>>>> I've been under the impression that you can set clock "flags" via the
>>>>>> device tree. Seems I need to re-check that ;)
>>>>>
>>>>> Right, you cannot set flags from the device tree. Also, setting these
>>>>> flags is done by the clock provider driver, not a consumer. Xen is the
>>>>> consumer.
>>>>
>>>>
>>>> Ok, thanks, then I think we can forget about using flags for the issue we
>>>> are
>>>> discussing here.
>>>>
>>>> Best regards
>>>>
>>>> Dirk
>>>>
>>>> P.S.: Would it be an option to merge the v4 patch we are discussing here,
>>>> then? From the discussion until here, it sounds to me that it's the best
>>>> option we have at the moment. Maybe improving it in the future, then.
>>>
>>> It might be a step in the right direction, but it doesn't really prevent
>>> clk_set_rate from changing properties of a clock owned by Xen.  This
>>> patch is incomplete.
>>
>>
>> Let me ask then: Do we have a practical example where it's not sufficient
>> practically?
>>
>> To my understanding, Xen people have been happy with the "clk_ignore_unused"
>> workaround since ~2 years, now [1]. And I think the "clk_ignore_unused"
>> workaround does mainly the same like the patch discussed here. It doesn't care
>> regarding clk_set_rate from changing properties, too?
>
> Let me premise that I appreciate what you are trying to achieve with
> this patch and I don't want to feature-creep it.
>
> However we are defining a new Device Tree binding,


I don't think so.

We are just using the existing one

https://git.kernel.org/cgit/linux/kernel/git/torvalds/linux.git/tree/Documentation/devicetree/bindings/clock/clock-bindings.txt#n66

, pick it from other device tree nodes (e.g. serial, timer etc) and 
add it to the hypervisor node. And then use this existing one with the 
existing well defined clock API.


> one which will have
> to be supported for a long time by both Xen and Linux, so at the very
> least we need to have the full picture. We need to understand if the
> binding if sufficient


Even if it's not sufficient, you can't change it.


> or if we need something different to solve the
> problem completely.


You might need anything additionally. E.g. an extension of the Linux 
kernel clock API to be able to modify the flags was proposed.

Best regards

Dirk

P.S.: I still would be interested if we do have a practical example 
where it's not sufficient practically?


> Once we understand that, I am happy to accept a partial implementation
> in Linux, as long as it is a step in the right direction. Does it make
> sense?
>
>
>
>> While I agree that the patch theoretically incomplete, if nobody has a real
>> world example I would think that from practical point of view it's sufficient
>> in a first step.
>>
>> If this is the case, I'd propose to fix the practical issue in a first step
>> with a patch (this one) which is sufficient to fix the issues the Xen users
>> have. And update the code for theoretical future issues in a second step.
>>
>> Best regards
>>
>> Dirk
>>
>> [1] http://bugs.xenproject.org/xen/bug/45
>>
>>
>>> We need to understand at least what it would take
>>> to have a complete solution.
>>>
>>> Michael, do you have any suggestions on how it would be possible to set
>>> CLK_SET_RATE_GATE and CLK_SET_PARENT_GATE for those clocks in a proper
>>> way?
>>>
>>> Like you wrote, I would imagine it needs to be done by the clock
>>> provider driver. Maybe to do that, it would be easier to have a new
>>> device tree property on the clock node, rather than listing phandle and
>>> clock-specifier pairs under the Xen node?
>>
>