Re: [PATCH v2 5/5] irqchip/irq-bcm7038-l1: Support brcm,int-fwd-mask

[Marc Zyngier <maz@kernel.org>]

On 23/09/2019 15:39, Florian Fainelli wrote:
> 
> 
> On 9/23/2019 1:52 AM, Marc Zyngier wrote:
>> On 22/09/2019 20:08, Florian Fainelli wrote:
>>>
>>>
>>> On 9/22/2019 5:38 AM, Marc Zyngier wrote:
>>>> On Fri, 13 Sep 2019 12:15:42 -0700
>>>> Florian Fainelli <f.fainelli@gmail.com> wrote:
>>>>
>>>>> On some specific chips like 7211 we need to leave some interrupts
>>>>> untouched/forwarded to the VPU which is another agent in the system
>>>>> making use of that interrupt controller hardware (goes to both ARM GIC
>>>>> and VPU L1 interrupt controller). Make that possible by using the
>>>>> existing brcm,int-fwd-mask property.
>>>>>
>>>>> Signed-off-by: Florian Fainelli <f.fainelli@gmail.com>
>>>>> ---
>>>>>  drivers/irqchip/irq-bcm7038-l1.c | 15 +++++++++++++--
>>>>>  1 file changed, 13 insertions(+), 2 deletions(-)
>>>>>
>>>>> diff --git a/drivers/irqchip/irq-bcm7038-l1.c b/drivers/irqchip/irq-bcm7038-l1.c
>>>>> index 0673a44bbdc2..811a34201dd4 100644
>>>>> --- a/drivers/irqchip/irq-bcm7038-l1.c
>>>>> +++ b/drivers/irqchip/irq-bcm7038-l1.c
>>>>> @@ -44,6 +44,7 @@ struct bcm7038_l1_chip {
>>>>>  	struct list_head	list;
>>>>>  	u32			wake_mask[MAX_WORDS];
>>>>>  #endif
>>>>> +	u32			irq_fwd_mask[MAX_WORDS];
>>>>>  	u8			affinity[MAX_WORDS * IRQS_PER_WORD];
>>>>>  };
>>>>>  
>>>>> @@ -265,6 +266,7 @@ static int __init bcm7038_l1_init_one(struct device_node *dn,
>>>>>  	resource_size_t sz;
>>>>>  	struct bcm7038_l1_cpu *cpu;
>>>>>  	unsigned int i, n_words, parent_irq;
>>>>> +	int ret;
>>>>>  
>>>>>  	if (of_address_to_resource(dn, idx, &res))
>>>>>  		return -EINVAL;
>>>>> @@ -278,6 +280,14 @@ static int __init bcm7038_l1_init_one(struct device_node *dn,
>>>>>  	else if (intc->n_words != n_words)
>>>>>  		return -EINVAL;
>>>>>  
>>>>> +	ret = of_property_read_u32_array(dn , "brcm,int-fwd-mask",
>>>>
>>>> What is the exact meaning of "fwd"? Forward? FirmWare Dementia?
>>>
>>> Here it is meant to be "forward", we have defined this property name
>>> before for irq-bcm7120-l2.c and felt like reusing the same name to avoid
>>> multiplying properties would be appropriate, see patch #4. If you prefer
>>> something named brcm,firmware-configured-mask, let me know.
>>
>> It's just a name, but I found it a bit confusing. Bah, never mind.
>>
>>>>
>>>>> +					 intc->irq_fwd_mask, n_words);
>>>>> +	if (ret != 0 && ret != -EINVAL) {
>>>>> +		/* property exists but has the wrong number of words */
>>>>> +		pr_err("invalid brcm,int-fwd-mask property\n");
>>>>> +		return -EINVAL;
>>>>> +	}
>>>>> +
>>>>>  	cpu = intc->cpus[idx] = kzalloc(sizeof(*cpu) + n_words * sizeof(u32),
>>>>>  					GFP_KERNEL);
>>>>>  	if (!cpu)
>>>>> @@ -288,8 +298,9 @@ static int __init bcm7038_l1_init_one(struct device_node *dn,
>>>>>  		return -ENOMEM;
>>>>>  
>>>>>  	for (i = 0; i < n_words; i++) {
>>>>> -		l1_writel(0xffffffff, cpu->map_base + reg_mask_set(intc, i));
>>>>> -		cpu->mask_cache[i] = 0xffffffff;
>>>>> +		l1_writel(0xffffffff & ~intc->irq_fwd_mask[i],
>>>>> +			  cpu->map_base + reg_mask_set(intc, i));
>>>>> +		cpu->mask_cache[i] = 0xffffffff & ~intc->irq_fwd_mask[i];
>>>>
>>>> I seem to remember that (0xffffffff & whatever) == whatever, as long as
>>>> 'whatever' is a 32bit quantity. So what it this for?
>>>
>>> It is 0xffff_ffff & ~whatever here.
>>
>> Which doesn't change anything.
>>
>>> In the absence of this property
>>> being specified, the data is all zeroed out, so we would have
>>> 0xffff_ffff & 0xffff_ffff which is 0xffff_ffff. If this property is
>>> specified, we would have one more or bits set, and it would be e.g.:
>>> 0x100 so we would have 0xffff_ffff & ~(0x100) = 0xffff_feff which is
>>> what we would want here to preserve whatever the firmware has already
>>> configured.
>>
>> OK, I must be stupid:
>>
>> #include <stdio.h>
>>
>> int main(int argc, char *argv[])
>> {
>> 	unsigned int v = 0x100;
>> 	printf ("%x\n", ~v);
>> }
>> maz@filthy-habit$ ./x
>> fffffeff
>>
>> You might as well OR it with zeroes, if you want.
> 
> Not sure I understand your point here.
> 
> We used to write 0xffff_ffff to both the hardware and the mask cache to
> have all interrupts masked by default. Now we want to have some bits
> optionally set to 0 (unmasked), based on the brcm,int-fwd-mask property,
> which is what this patch achieves (or tries to). If we write, say
> 0xffff_feff to the hardware, which has a Write Only register behavior,
> then we also want to have the mask cache be set to the same value for
> consistency if nothing else. Am I failing at doing what I just described
> and also failing at see it?

You write this:

>  	for (i = 0; i < n_words; i++) {
> -		l1_writel(0xffffffff, cpu->map_base + reg_mask_set(intc, i));
> -		cpu->mask_cache[i] = 0xffffffff;
> +		l1_writel(0xffffffff & ~intc->irq_fwd_mask[i],
> +			  cpu->map_base + reg_mask_set(intc, i));
> +		cpu->mask_cache[i] = 0xffffffff & ~intc->irq_fwd_mask[i];
>  	}

And I'm saying that this is strictly equivalent to:

 	for (i = 0; i < n_words; i++) {
		l1_writel(~intc->irq_fwd_mask[i],
			  cpu->map_base + reg_mask_set(intc, i));
		cpu->mask_cache[i] = ~intc->irq_fwd_mask[i];
 	}

without this 0xffffffff that does exactly nothing (I'm pretty sure the
compiler drops it anyway).

	M.
-- 
Jazz is not dead, it just smells funny...