[PATCH] hw/core/clock: always iterate through childs in clock_propagate_period

[PATCH] hw/core/clock: always iterate through childs in clock_propagate_period

[Raphael Poggi]

When dealing with few clocks depending with each others, sometimes
we might only want to update the multiplier/diviser on a specific clock
(cf clockB in drawing below) and call "clock_propagate(clockA)" to
update the childs period according to the potential new multiplier/diviser values.

+--------+     +--------+      +--------+
| clockA | --> | clockB |  --> | clockC |
+--------+     +--------+      +--------+

The actual code would not allow that because, since we cannot call
"clock_propagate" directly on a child, it would exit on the
first child has the period has not changed for clockB, only clockC is
impacted in our example.

Signed-off-by: Raphael Poggi <raphael.poggi@lynxleap.co.uk>
---
 hw/core/clock.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/hw/core/clock.c b/hw/core/clock.c
index a19c7db7df..85421f8b55 100644
--- a/hw/core/clock.c
+++ b/hw/core/clock.c
@@ -101,8 +101,9 @@ static void clock_propagate_period(Clock *clk, bool call_callbacks)
             if (call_callbacks) {
                 clock_call_callback(child, ClockUpdate);
             }
-            clock_propagate_period(child, call_callbacks);
         }
+
+        clock_propagate_period(child, call_callbacks);
     }
 }
 
-- 
2.44.0

Re: [PATCH] hw/core/clock: always iterate through childs in clock_propagate_period

[Philippe Mathieu-Daudé]

Hi Raphael,

On 18/4/24 21:16, Raphael Poggi wrote:
> When dealing with few clocks depending with each others, sometimes
> we might only want to update the multiplier/diviser on a specific clock
> (cf clockB in drawing below) and call "clock_propagate(clockA)" to
> update the childs period according to the potential new multiplier/diviser values.
> 
> +--------+     +--------+      +--------+
> | clockA | --> | clockB |  --> | clockC |
> +--------+     +--------+      +--------+
> 
> The actual code would not allow that because, since we cannot call
> "clock_propagate" directly on a child, it would exit on the
> first child has the period has not changed for clockB, only clockC is

Typo "as the period has not changed"?

Why can't you call clock_propagate() on a child?

> impacted in our example.
> 
> Signed-off-by: Raphael Poggi <raphael.poggi@lynxleap.co.uk>
> ---
>   hw/core/clock.c | 3 ++-
>   1 file changed, 2 insertions(+), 1 deletion(-)
> 
> diff --git a/hw/core/clock.c b/hw/core/clock.c
> index a19c7db7df..85421f8b55 100644
> --- a/hw/core/clock.c
> +++ b/hw/core/clock.c
> @@ -101,8 +101,9 @@ static void clock_propagate_period(Clock *clk, bool call_callbacks)
>               if (call_callbacks) {
>                   clock_call_callback(child, ClockUpdate);
>               }
> -            clock_propagate_period(child, call_callbacks);
>           }
> +
> +        clock_propagate_period(child, call_callbacks);
>       }
>   }
>   

Re: [PATCH] hw/core/clock: always iterate through childs in clock_propagate_period

[Raphael Poggi]

Hi Philippe,

Le jeu. 18 avr. 2024 à 20:43, Philippe Mathieu-Daudé
<philmd@linaro.org> a écrit :
>
> Hi Raphael,
>
> On 18/4/24 21:16, Raphael Poggi wrote:
> > When dealing with few clocks depending with each others, sometimes
> > we might only want to update the multiplier/diviser on a specific clock
> > (cf clockB in drawing below) and call "clock_propagate(clockA)" to
> > update the childs period according to the potential new multiplier/diviser values.
> >
> > +--------+     +--------+      +--------+
> > | clockA | --> | clockB |  --> | clockC |
> > +--------+     +--------+      +--------+
> >
> > The actual code would not allow that because, since we cannot call
> > "clock_propagate" directly on a child, it would exit on the
> > first child has the period has not changed for clockB, only clockC is
>
> Typo "as the period has not changed"?

That's a typo indeed, thanks!

>
> Why can't you call clock_propagate() on a child?

There is an assert "assert(clk->source == NULL);" in clock_propagate().
If I am not wrong, clk->source is set when the clock has a parent.

>
> > impacted in our example.
> >
> > Signed-off-by: Raphael Poggi <raphael.poggi@lynxleap.co.uk>
> > ---
> >   hw/core/clock.c | 3 ++-
> >   1 file changed, 2 insertions(+), 1 deletion(-)
> >
> > diff --git a/hw/core/clock.c b/hw/core/clock.c
> > index a19c7db7df..85421f8b55 100644
> > --- a/hw/core/clock.c
> > +++ b/hw/core/clock.c
> > @@ -101,8 +101,9 @@ static void clock_propagate_period(Clock *clk, bool call_callbacks)
> >               if (call_callbacks) {
> >                   clock_call_callback(child, ClockUpdate);
> >               }
> > -            clock_propagate_period(child, call_callbacks);
> >           }
> > +
> > +        clock_propagate_period(child, call_callbacks);
> >       }
> >   }
> >
>

Re: [PATCH] hw/core/clock: always iterate through childs in clock_propagate_period

[Peter Maydell]

On Thu, 18 Apr 2024 at 21:39, Raphael Poggi
<raphael.poggi@lynxleap.co.uk> wrote:
>
> Hi Philippe,
>
> Le jeu. 18 avr. 2024 à 20:43, Philippe Mathieu-Daudé
> <philmd@linaro.org> a écrit :
> >
> > Hi Raphael,
> >
> > On 18/4/24 21:16, Raphael Poggi wrote:
> > > When dealing with few clocks depending with each others, sometimes
> > > we might only want to update the multiplier/diviser on a specific clock
> > > (cf clockB in drawing below) and call "clock_propagate(clockA)" to
> > > update the childs period according to the potential new multiplier/diviser values.
> > >
> > > +--------+     +--------+      +--------+
> > > | clockA | --> | clockB |  --> | clockC |
> > > +--------+     +--------+      +--------+
> > >
> > > The actual code would not allow that because, since we cannot call
> > > "clock_propagate" directly on a child, it would exit on the
> > > first child has the period has not changed for clockB, only clockC is
> >
> > Typo "as the period has not changed"?
>
> That's a typo indeed, thanks!
>
> >
> > Why can't you call clock_propagate() on a child?
>
> There is an assert "assert(clk->source == NULL);" in clock_propagate().
> If I am not wrong, clk->source is set when the clock has a parent.

I think that assertion is probably there because we didn't
originally have the idea of a clock having a multiplier/divider
setting. So the idea was that calling clock_propagate() on a
clock with a parent would always be wrong, because the only
reason for its period to change would be if the parent had
changed, and if the parent changes then clock_propagate()
should be called on the parent.

We added mul/div later, and we (I) didn't think through all
the consequences. If you change the mul/div settings on
clockB in this example then you need to call clock_propagate()
on it, so we should remove that assert(). Then when you change
the mul/div on clockB you can directly clock_propagate(clockB),
and I don't think you need this patch at that point.

thanks
-- PMM

Re: [PATCH] hw/core/clock: always iterate through childs in clock_propagate_period

[Raphael Poggi]

Hi Peter,

Le ven. 19 avr. 2024 à 16:08, Peter Maydell <peter.maydell@linaro.org> a écrit :
>
> On Thu, 18 Apr 2024 at 21:39, Raphael Poggi
> <raphael.poggi@lynxleap.co.uk> wrote:
> >
> > Hi Philippe,
> >
> > Le jeu. 18 avr. 2024 à 20:43, Philippe Mathieu-Daudé
> > <philmd@linaro.org> a écrit :
> > >
> > > Hi Raphael,
> > >
> > > On 18/4/24 21:16, Raphael Poggi wrote:
> > > > When dealing with few clocks depending with each others, sometimes
> > > > we might only want to update the multiplier/diviser on a specific clock
> > > > (cf clockB in drawing below) and call "clock_propagate(clockA)" to
> > > > update the childs period according to the potential new multiplier/diviser values.
> > > >
> > > > +--------+     +--------+      +--------+
> > > > | clockA | --> | clockB |  --> | clockC |
> > > > +--------+     +--------+      +--------+
> > > >
> > > > The actual code would not allow that because, since we cannot call
> > > > "clock_propagate" directly on a child, it would exit on the
> > > > first child has the period has not changed for clockB, only clockC is
> > >
> > > Typo "as the period has not changed"?
> >
> > That's a typo indeed, thanks!
> >
> > >
> > > Why can't you call clock_propagate() on a child?
> >
> > There is an assert "assert(clk->source == NULL);" in clock_propagate().
> > If I am not wrong, clk->source is set when the clock has a parent.
>
> I think that assertion is probably there because we didn't
> originally have the idea of a clock having a multiplier/divider
> setting. So the idea was that calling clock_propagate() on a
> clock with a parent would always be wrong, because the only
> reason for its period to change would be if the parent had
> changed, and if the parent changes then clock_propagate()
> should be called on the parent.
>
> We added mul/div later, and we (I) didn't think through all
> the consequences. If you change the mul/div settings on
> clockB in this example then you need to call clock_propagate()
> on it, so we should remove that assert(). Then when you change
> the mul/div on clockB you can directly clock_propagate(clockB),
> and I don't think you need this patch at that point.

Alright, that makes sense, is that OK if I send a patch removing the assert ?

Thanks,
>
> thanks
> -- PMM

Re: [PATCH] hw/core/clock: always iterate through childs in clock_propagate_period

[Peter Maydell]

On Fri, 19 Apr 2024 at 17:09, Raphael Poggi
<raphael.poggi@lynxleap.co.uk> wrote:
>
> Hi Peter,
>
> Le ven. 19 avr. 2024 à 16:08, Peter Maydell <peter.maydell@linaro.org> a écrit :
> >
> > On Thu, 18 Apr 2024 at 21:39, Raphael Poggi
> > <raphael.poggi@lynxleap.co.uk> wrote:
> > > There is an assert "assert(clk->source == NULL);" in clock_propagate().
> > > If I am not wrong, clk->source is set when the clock has a parent.
> >
> > I think that assertion is probably there because we didn't
> > originally have the idea of a clock having a multiplier/divider
> > setting. So the idea was that calling clock_propagate() on a
> > clock with a parent would always be wrong, because the only
> > reason for its period to change would be if the parent had
> > changed, and if the parent changes then clock_propagate()
> > should be called on the parent.
> >
> > We added mul/div later, and we (I) didn't think through all
> > the consequences. If you change the mul/div settings on
> > clockB in this example then you need to call clock_propagate()
> > on it, so we should remove that assert(). Then when you change
> > the mul/div on clockB you can directly clock_propagate(clockB),
> > and I don't think you need this patch at that point.
>
> Alright, that makes sense, is that OK if I send a patch removing the assert ?

Yes, please do.

-- PMM

Re: [PATCH] hw/core/clock: always iterate through childs in clock_propagate_period

[Philippe Mathieu-Daudé]

On 19/4/24 18:08, Raphael Poggi wrote:
> Hi Peter,
> 
> Le ven. 19 avr. 2024 à 16:08, Peter Maydell <peter.maydell@linaro.org> a écrit :
>>
>> On Thu, 18 Apr 2024 at 21:39, Raphael Poggi
>> <raphael.poggi@lynxleap.co.uk> wrote:
>>>
>>> Hi Philippe,
>>>
>>> Le jeu. 18 avr. 2024 à 20:43, Philippe Mathieu-Daudé
>>> <philmd@linaro.org> a écrit :
>>>>
>>>> Hi Raphael,
>>>>
>>>> On 18/4/24 21:16, Raphael Poggi wrote:
>>>>> When dealing with few clocks depending with each others, sometimes
>>>>> we might only want to update the multiplier/diviser on a specific clock
>>>>> (cf clockB in drawing below) and call "clock_propagate(clockA)" to
>>>>> update the childs period according to the potential new multiplier/diviser values.
>>>>>
>>>>> +--------+     +--------+      +--------+
>>>>> | clockA | --> | clockB |  --> | clockC |
>>>>> +--------+     +--------+      +--------+
>>>>>
>>>>> The actual code would not allow that because, since we cannot call
>>>>> "clock_propagate" directly on a child, it would exit on the
>>>>> first child has the period has not changed for clockB, only clockC is
>>>>
>>>> Typo "as the period has not changed"?
>>>
>>> That's a typo indeed, thanks!
>>>
>>>>
>>>> Why can't you call clock_propagate() on a child?
>>>
>>> There is an assert "assert(clk->source == NULL);" in clock_propagate().
>>> If I am not wrong, clk->source is set when the clock has a parent.
>>
>> I think that assertion is probably there because we didn't
>> originally have the idea of a clock having a multiplier/divider
>> setting. So the idea was that calling clock_propagate() on a
>> clock with a parent would always be wrong, because the only
>> reason for its period to change would be if the parent had
>> changed, and if the parent changes then clock_propagate()
>> should be called on the parent.
>>
>> We added mul/div later, and we (I) didn't think through all
>> the consequences. If you change the mul/div settings on
>> clockB in this example then you need to call clock_propagate()
>> on it, so we should remove that assert(). Then when you change
>> the mul/div on clockB you can directly clock_propagate(clockB),
>> and I don't think you need this patch at that point.
> 
> Alright, that makes sense, is that OK if I send a patch removing the assert ?

Sure, that is welcomed :)

Regards,

Phil.