[linux-sunxi] Re: [PATCH v3] dt-bindings: Add a clocks property to the simple-framebuffer binding

[Hans de Goede]

Hi,

On 10/04/2014 12:56 AM, jonsmirl at gmail.com wrote:
> On Fri, Oct 3, 2014 at 4:08 PM, Rob Herring <robherring2 at gmail.com> wrote:
>> On Fri, Oct 3, 2014 at 12:34 PM, Hans de Goede <hdegoede at redhat.com> wrote:
>>> Hi,
>>>
>>> On 10/03/2014 05:57 PM, Rob Herring wrote:
>>>> On Fri, Oct 3, 2014 at 9:05 AM, Hans de Goede <hdegoede at redhat.com> wrote:
>>>>> A simple-framebuffer node represents a framebuffer setup by the firmware /
>>>>> bootloader. Such a framebuffer may have a number of clocks in use, add a
>>>>> property to communicate this to the OS.
>>>>>
>>>>> Signed-off-by: Hans de Goede <hdegoede at redhat.com>
>>>>> Reviewed-by: Mike Turquette <mturquette at linaro.org>
>>>>>
>>>>> --
>>>>> Changes in v2:
>>>>> -Added Reviewed-by: Mike Turquette <mturquette at linaro.org>
>>>>> Changes in v3:
>>>>> -Updated description to make clear simplefb deals with more then just memory
>>>>
>>>> NAK. "Fixing" the description is not what I meant and does not address
>>>> my concerns. Currently, simplefb is configuration data. It is
>>>> auxiliary data about how a chunk of memory is used. Using it or not
>>>> has no side effects on the hardware setup, but you are changing that
>>>> aspect. You are mixing in a hardware description that is simply
>>>> inaccurate.
>>>
>>> Memory is hardware too, what simplefb is is best seen as a virtual device, and
>>> even virtual devices have hardware resources they need, such as a chunk of memory,
>>> which the kernel should not touch other then use it as a framebuffer, likewise
>>> on some systems the virtual device needs clocks to keep running.
>>>
>>>> The kernel has made the decision to turn off "unused" clocks. If its
>>>> determination of what is unused is wrong, then it is not a problem to
>>>> fix in DT.
>>>
>>> No, it is up to DT to tell the kernel what clocks are used, that is how it works
>>> for any other device. I don't understand why some people keep insisting simplefb
>>> for some reason is o so very very special, because it is not special, it needs
>>> resources, and it needs to tell the kernel about this or bad things happen.
>>
>> No, the DT describes the connections of clocks from h/w block to h/w
>> block. Their use is implied by the connection.
>>
>> And yes, as the other thread mentioned DT is more than just hardware
>> information. However, what you are adding IS hardware information and
>> clearly has a place somewhere else. And adding anything which is not
>> hardware description gets much more scrutiny.
>>
>>> More over it is a bit late to start making objections now. This has already been
>>> discussed to death for weeks now, and every argument against this patch has already
>>> been countered multiple times (including the one you are making now). Feel free to
>>> read the entire thread in the archives under the subject:
>>> "[PATCH 4/4] simplefb: add clock handling code"
>>
>> You are on v2 and I hardly see any consensus on the v1 thread. Others
>> have made suggestions which I would agree with and you've basically
>> ignored them.
>>
>>> At one point in time we need to stop bickering and make a decision, that time has
>>> come now, so please lets get this discussion over with and merge this, so that
>>> we can all move on and spend out time in a more productive manner.
>>
>> Not an effective argument to get things merged.
> 
> If there is not good solution to deferring clock clean up in the
> kernel, another approach is to change how simple-framebuffer is
> described in the device tree....
> 
> Right now it is a stand-alone item that looks like a device node, but
> it isn't a device.
> 
> framebuffer {
>     compatible = "simple-framebuffer";
>     reg = <0x1d385000 (1600 * 1200 * 2)>;
>     width = <1600>;
>     height = <1200>;
>     stride = <(1600 * 2)>;
>     format = "r5g6b5";
> };
> 
> How about something like this?
> 
> reserved-memory {
>     #address-cells = <1>;
>     #size-cells = <1>;
>     ranges;
> 
>     display_reserved: framebuffer at 78000000 {
>         reg = <0x78000000  (1600 * 1200 * 2)>;
>     };
> };
> 
> lcd0: lcd-controller at 820000 {
>     compatible = "marvell,dove-lcd";
>     reg = <0x820000 0x1000>;
>     interrupts = <47>;
>     clocks = <&si5351 0>;
>     clock-names = "ext_ref_clk_1";
> };
> 
> chosen {
>     boot-framebuffer {
>        compatible = "simple-framebuffer";
>        device = <&lcd0>;
>        framebuffer = <&display_reserved>;
>        width = <1600>;
>        height = <1200>;
>        stride = <(1600 * 2)>;
>        format = "r5g6b5";
>     };
> }
> 
> 
> This moves the definition of the boot framebuffer setup into the area
> where bootloader info is suppose to go. Then you can use the phandle
> to follow the actual device chains and protect the clocks and
> regulators. To make that work you are required to provide an accurate
> description of the real video hardware so that this chain can be
> followed.

This will not work, first of all multiple blocks may be involved, so
the device = in the boot-framebuffer would need to be a list. That in
itself is not a problem, the problem is that the blocks used may have
multiple clocks, of which the setup mode likely uses only a few.

So if we do things this way, we end up keeping way to many clocks
enabled.

This does nicely show why the clocks really belong in the simplefb
node though. What you suggest above, would lead to simplefb having
access to the hardware info / description of the display engine
blocks, putting the clocks info in the display engine nodes, so
keeping hardware description with hardware description as people
want.

But as said that does not tell the kernel which clocks are actually
used for the setup mode, so it does not provide the info the kernel
needs. What the kernel needs is not hardware description, but a
list of clocks which are *actually used* by the setup mode.

Thinking more about this the clocks property in the simplefb node
is just like the reg property. Both are properties normally only
found in real harware nodes, not in an informational node like
simplefb.

But for the kernel to be able to actually use the simplefb it
needs to know which memory is used by the framebuffer. One could
argue that a reg property is hardware description, and thus does not
belong in the simplefb node, and that the kernel should just magically
figure out which memory is used.

Everyone sees that the kernel cannot magically figure out which memory
is used by the simplefb. Yet people expect the kernel to somehow
magically figure out which clocks are used, to avoid accidentally turning
them off.

This is not consistent, either the kernel needs to be told these things,
and then it needs to be told all of them, or hardware node properties
like a reg property do not belong in an informational node, and then the
reg property should not be in the simplefb node either, and the kernel
should somehow magically figure out where the memory lives, just like
some people are expecting the kernel to magically figure out which
clocks are used.

Regards,

Hans