Re: [PATCH v1 2/2] arch/arm: time: Add support for parsing interrupts by names

[Michal Orzel <michal.orzel@xxxxxxx>]

Hi Bertrand,

On 09/03/2023 13:19, Bertrand Marquis wrote:
> 
> 
> Hi Michal,
> 
>> On 9 Mar 2023, at 12:35, Michal Orzel <michal.orzel@xxxxxxx> wrote:
>>
>>
>>
>> On 09/03/2023 11:39, Bertrand Marquis wrote:
>>>
>>>
>>> Hi Michal,
>>>
>>>> On 9 Mar 2023, at 11:05, Michal Orzel <michal.orzel@xxxxxxx> wrote:
>>>>
>>>>
>>>>
>>>> On 09/03/2023 09:02, Bertrand Marquis wrote:
>>>>>
>>>>>
>>>>> Hi Stefano,
>>>>>
>>>>>> On 7 Mar 2023, at 22:02, Stefano Stabellini <sstabellini@xxxxxxxxxx> 
>>>>>> wrote:
>>>>>>
>>>>>> On Tue, 7 Mar 2023, Bertrand Marquis wrote:
>>>>>>>> On 7 Mar 2023, at 11:09, Andrei Cherechesu (OSS) 
>>>>>>>> <andrei.cherechesu@xxxxxxxxxxx> wrote:
>>>>>>>>
>>>>>>>> From: Andrei Cherechesu <andrei.cherechesu@xxxxxxx>
>>>>>>>>
>>>>>>>> Added support for parsing the ARM generic timer interrupts DT
>>>>>>>> node by the "interrupt-names" property, if it is available.
>>>>>>>>
>>>>>>>> If not available, the usual parsing based on the expected
>>>>>>>> IRQ order is performed.
>>>>>>>>
>>>>>>>> Also added the "hyp-virt" PPI to the timer PPI list, even
>>>>>>>> though it's currently not in use. If the "hyp-virt" PPI is
>>>>>>>> not found, the hypervisor won't panic.
>>>>>>>>
>>>>>>>> Signed-off-by: Andrei Cherechesu <andrei.cherechesu@xxxxxxx>
>>>>>>>> ---
>>>>>>>> xen/arch/arm/include/asm/time.h |  3 ++-
>>>>>>>> xen/arch/arm/time.c             | 26 ++++++++++++++++++++++----
>>>>>>>> 2 files changed, 24 insertions(+), 5 deletions(-)
>>>>>>>>
>>>>>>>> diff --git a/xen/arch/arm/include/asm/time.h 
>>>>>>>> b/xen/arch/arm/include/asm/time.h
>>>>>>>> index 4b401c1110..49ad8c1a6d 100644
>>>>>>>> --- a/xen/arch/arm/include/asm/time.h
>>>>>>>> +++ b/xen/arch/arm/include/asm/time.h
>>>>>>>> @@ -82,7 +82,8 @@ enum timer_ppi
>>>>>>>>  TIMER_PHYS_NONSECURE_PPI = 1,
>>>>>>>>  TIMER_VIRT_PPI = 2,
>>>>>>>>  TIMER_HYP_PPI = 3,
>>>>>>>> -    MAX_TIMER_PPI = 4,
>>>>>>>> +    TIMER_HYP_VIRT_PPI = 4,
>>>>>>>> +    MAX_TIMER_PPI = 5,
>>>>>>>> };
>>>>>>>>
>>>>>>>> /*
>>>>>>>> diff --git a/xen/arch/arm/time.c b/xen/arch/arm/time.c
>>>>>>>> index 433d7be909..794da646d6 100644
>>>>>>>> --- a/xen/arch/arm/time.c
>>>>>>>> +++ b/xen/arch/arm/time.c
>>>>>>>> @@ -38,6 +38,14 @@ uint32_t __read_mostly timer_dt_clock_frequency;
>>>>>>>>
>>>>>>>> static unsigned int timer_irq[MAX_TIMER_PPI];
>>>>>>>>
>>>>>>>> +static const char *timer_irq_names[MAX_TIMER_PPI] = {
>>>>>>>> +    [TIMER_PHYS_SECURE_PPI] = "sec-phys",
>>>>>>>> +    [TIMER_PHYS_NONSECURE_PPI] = "phys",
>>>>>>>> +    [TIMER_VIRT_PPI] = "virt",
>>>>>>>> +    [TIMER_HYP_PPI] = "hyp-phys",
>>>>>>>> +    [TIMER_HYP_VIRT_PPI] = "hyp-virt",
>>>>>>>> +};
>>>>>>>> +
>>>>>>>
>>>>>>> I would need some reference or a pointer to some doc to check those.
>>>>>>>
>>>>>>>> unsigned int timer_get_irq(enum timer_ppi ppi)
>>>>>>>> {
>>>>>>>>  ASSERT(ppi >= TIMER_PHYS_SECURE_PPI && ppi < MAX_TIMER_PPI);
>>>>>>>> @@ -149,15 +157,25 @@ static void __init init_dt_xen_time(void)
>>>>>>>> {
>>>>>>>>  int res;
>>>>>>>>  unsigned int i;
>>>>>>>> +    bool has_names;
>>>>>>>> +
>>>>>>>> +    has_names = dt_property_read_bool(timer, "interrupt-names");
>>>>>>>>
>>>>>>>>  /* Retrieve all IRQs for the timer */
>>>>>>>>  for ( i = TIMER_PHYS_SECURE_PPI; i < MAX_TIMER_PPI; i++ )
>>>>>>>>  {
>>>>>>>> -        res = platform_get_irq(timer, i);
>>>>>>>> -
>>>>>>>> -        if ( res < 0 )
>>>>>>>> +        if ( has_names )
>>>>>>>> +            res = platform_get_irq_byname(timer, timer_irq_names[i]);
>>>>>>>> +        else
>>>>>>>> +            res = platform_get_irq(timer, i);
>>>>>>>> +
>>>>>>>> +        if ( res > 0 )
>>>>>>>
>>>>>>> The behaviour of the code is changed here compared to the current
>>>>>>> version as res = 0 will now generate a panic.
>>>>>>>
>>>>>>> Some device tree might not specify an interrupt number and just put
>>>>>>> 0 and Xen will now panic on those systems.
>>>>>>> As I have no idea if such systems exists and the behaviour is modified
>>>>>>> you should justify this and mention it in the commit message or keep
>>>>>>> the old behaviour and let 0 go through without a panic.
>>>>>>>
>>>>>>> @stefano, julien any idea here ? should just keep the old behaviour ?
>>>>>>
>>>>>> platform_get_irq returns 0 if the irq is 0. The irq cannot be 0 because
>>>>>> 0 is reserved for SGIs, not PPIs. So I think it is OK to consider 0 an
>>>>>> error.
>>>>>
>>>>> Problem here is that a DTB might not specify all interrupts and just put
>>>>> 0 for the one not used (or not available for example if you have no secure
>>>>> world).
>>>> Xen requires presence of EL3,EL2 and on such system, at least the 
>>>> following timers needs to be there
>>>> according to Arm ARM:
>>>> - EL3 phys (if EL3 is there)
>>>
>>> This might be needed by EL3 but not by Xen.
>> Xen requires system with EL3 and if there is EL3, both Arm spec and dt 
>> bindings requires sec-phys timer to be there.
>> So it would be very strange to see a fake interrupt with IRQ being 0. But if 
>> we relly want to only care about
>> what Xen needs, then we could live with that (although it is difficult for 
>> me to find justification for 0 there).
>> Device trees are created per system and if system has EL3, then why forcing 
>> 0 to be listed for sec-phys timer?
>>
> 
> Let's see that on the other angle: why should Xen check stuff that it does 
> not need ?
Because apart from what it needs or not, there is a matter of a failure in Xen.
Xen exposes timer IRQs to dom0 as they were taken from dtb and allowing 0 for 
any of the timer IRQ would result
in a Xen failure when reserving such IRQ. Xen presets SGI IRQs, meaning bits 
0:15 in allocated_irqs bitmap are set.
This is why when calling vgic_reserve_virq and passing SGI always results in 
calling a BUG().

So we have two options:
- panic earlier with a meaningful message when IRQ is 0
- let Xen continue and reach BUG which would be not that obvious for people 
using but not knowing Xen

I think first option is always better.

~Michal