Re: Switching from xenomai 3.2 IPIPE to xenomai 3.3 Dovetail

[Philippe Gerum <rpm@xenomai.org>]

Philippe Gerum <rpm@xenomai.org> writes:

> François Legal <devel@thom.fr.eu.org> writes:
>
>> Le Mardi, Novembre 19, 2024 11:23 CET, Philippe Gerum <rpm@xenomai.org> a écrit: 
>>  
>>> François Legal <francois.legal@thom.fr.eu.org> writes:
>>> 
>>> > Le Mardi, Novembre 19, 2024 10:26 CET, Julien Aube <julien.aube@smile.fr> a écrit: 
>>> >  
>>> >> Hello,
>>> >> 
>>> >> Reading this exchange it reminds me of a problem I had a few months ago 
>>> >> on the same platform.
>>> >> 
>>> >> Here is the thead :
>>> >> 
>>> >> https://lore.kernel.org/xenomai/6814fa90-015e-91df-6941-db643e4d4229@smile.fr/#r
>>> >> 
>>> >> 
>>> >> Conclusion was that disabling the scutimer(*) on DTS level was enough to 
>>> >> restore the correct usage and I did not found
>>> >> any bad behaviour after that. The product has passed all the rt-test 
>>> >> afterward.
>>> >> 
>>> >> 
>>> >> (*)
>>> >> &scutimer {
>>> >>      /* scutimer is forced to disabled since it does not work with 
>>> >> Xenomai */
>>> >>      status = "disabled";
>>> >> };
>>> >> 
>>> >> Maybe i'm completely off the subject, but maybe it's worth having a look ?
>>> >> 
>>> >> Julien
>>> >> 
>>> >> 
>>> >
>>> > Thanks. That fixes the problem. I'll dig into that driver to try to fix it and submit a patch then.
>>> >
>>> 
>>> This may help:
>>> 
>>> diff --git a/arch/arm/kernel/smp_twd.c b/arch/arm/kernel/smp_twd.c
>>> index 8377f1d030c3a..a28f6b7f7eabc 100644
>>> --- a/arch/arm/kernel/smp_twd.c
>>> +++ b/arch/arm/kernel/smp_twd.c
>>> @@ -181,7 +181,7 @@ static irqreturn_t twd_handler(int irq, void *dev_id)
>>>  {
>>>  	struct clock_event_device *evt = dev_id;
>>>  
>>> -	if (twd_timer_ack()) {
>>> +	if (!clockevent_is_oob(evt) || twd_timer_ack()) {
>>>  		clockevents_handle_event(evt);
>>>  		return IRQ_HANDLED;
>>>  	}
>>> 
>>
>> No noticeable effect here. I tried to dig a little more into this, but I can't find anything.
>> What I can tell, is that the TWD interrupts triggers on both cores when hung at boot. TWD registers settings seems OK in terms of prescaler and everything. By checking with JTAG debugger, I can see that xnintr_core_clock_handler is being called in the TWD service routine (twd_handler).
>>
>> I'm trying a little harder, but don't really know from which angle !
>
> The issue is as follows:
>
> When the proxy tick hijacks the timer to hand over its control to
> Xenomai, twd_handler() may be called either from the oob stage upon a
> real tick coming from the hardware, _and_ from the in-band stage as a
> consequence of emulating a common kernel tick, see [1] in the "Under the
> hood" section.
>
> The way twd_handler() operates ATM causes the hardware to be sent an ack
> on every invocation, which is wrong wrt the in-band tick emulation,
> which is most likely wrecking the logic afterwards.
>
> [1] https://v4.xenomai.org/dovetail/porting/timer/#proxy-tick-logic

diff --git a/arch/arm/kernel/smp_twd.c b/arch/arm/kernel/smp_twd.c
index 8377f1d030c3a..194d3cc7ca2c8 100644
--- a/arch/arm/kernel/smp_twd.c
+++ b/arch/arm/kernel/smp_twd.c
@@ -181,7 +181,7 @@ static irqreturn_t twd_handler(int irq, void *dev_id)
 {
 	struct clock_event_device *evt = dev_id;
 
-	if (twd_timer_ack()) {
+	if ((running_inband() && clockevent_is_oob(evt)) || twd_timer_ack()) {
 		clockevents_handle_event(evt);
 		return IRQ_HANDLED;
 	}

-- 
Philippe.