On 12/12/2021 23.37, Marc Zyngier wrote:
>> @@ -75,6 +75,7 @@
>> #define AIC_EVENT_TYPE GENMASK(31, 16)
>> #define AIC_EVENT_NUM GENMASK(15, 0)
>>
>> +#define AIC_EVENT_TYPE_FIQ 0 /* Software use */
>
> What does 'SW use' mean? Are you using the fact that the event
> register never returns 0 in the top bits?
Yeah. Since we're switching to tree IRQs we can use the raw hardware
event as the hwirq, and save some cycles munging fields, but we still
need a place for FIQ hwirq numbers to live. Zero here means "no
IRQ/spurious" to the hardware, so it's a convenient place to stick FIQs.
Note that the top-level IRQ handler function does check this field, so
even if newer hardware starts doing something silly here (which I highly
doubt...) it would never end up forwarded to the IRQ domain code without
further code changes.
>> - aic_ic_write(ic, AIC_MASK_SET + MASK_REG(irqd_to_hwirq(d)),
>> - MASK_BIT(irqd_to_hwirq(d)));
>> + u32 irq = FIELD_GET(AIC_EVENT_NUM, hwirq);
>
> This expression is used quite a few times, and could use a helper
> clarifying its purpose (converting the event/hwirq to an index?).
> 'irq' is a bit of a misnomer too, but I struggle to find another
> name...
Ack, I'll add a helper. It's extracting the IRQ number field from the
hwirq number. This is relatively trivial at this point in the patch set
(where the only other field is the constant type = 1), but it makes more
sense once I add the die field later on.
>> @@ -492,13 +497,13 @@ static struct irq_chip fiq_chip = {
>> static int aic_irq_domain_map(struct irq_domain *id, unsigned int irq,
>> irq_hw_number_t hw)
>> {
>> - struct aic_irq_chip *ic = id->host_data;
>> + u32 type = FIELD_GET(AIC_EVENT_TYPE, hw);
>>
>> - if (hw < ic->nr_hw) {
>> + if (type == AIC_EVENT_TYPE_HW) {
>> irq_domain_set_info(id, irq, hw, &aic_chip, id->host_data,
>> handle_fasteoi_irq, NULL, NULL);
>> irqd_set_single_target(irq_desc_get_irq_data(irq_to_desc(irq)));
>> - } else {
>> + } else if (type == AIC_EVENT_TYPE_FIQ) {
>
> Do we need to check for FIQ? This should be the case by construction,
> right? If there is a risk that it isn't the case, then we probably
> need a default case (and the whole thing would be better written as a
> switch() statement).
Yes, it should be the case by construction; this can just be an else.
I'll change it.
--
Hector Martin (marcan@xxxxxxxxx)
Public Key: https://mrcn.st/pub
