On Thu, Feb 14, 2019 at 4:13 AM Suman Anna <s-anna@xxxxxx> wrote: > [Me] > > To be able to use hierarchical interrupt domain in the kernel, the top > > interrupt controller must use the hierarchical (v2) irqdomain, so > > if this is anything else than the ARM GIC it will be an interesting > > undertaking to handle this. > > These are interrupt lines coming towards the host processor running > Linux and are directly connected to the ARM GIC. This INTC module is > actually an PRUSS internal interrupt controller that can take in 64 (on > most SoCs) external events/interrupt sources and multiplexing them > through two layers of many-to-one events-to-intr channels & > intr-channels-to-host interrupts. Couple of the host interrupts go to > the PRU cores themselves while the remaining ones come out of the IP to > connect to other GICs in the SoC. If the muxing is static (like set up once at probe) so that while the system is running, there is one and one only event mapped to the GIC from the component below it, then it is hierarchical. > We have implemented this as an irqchip using chained interrupt handlers > with the consumers using the event numbers on the Linux-side. The PRUs > also access some of the associated registers for clearing an event source. Chaining with cascading is when two or more interrupts fire the same upper level (say GIC) IRQ. If there is a 1:1 mapping, it is not chained/cascaded but hierarchical. I understand you used old irqdomain/chip frameworks in the past, because everyone was working around the fact that they didn't have an abstraction for hierarchical IRQs. Using chained interrupts and custom 1:1 maps and assigning a long list of IRQs like this patch does was the most common workaround. But we should step out of that habit now. Different levels of the IRQ handling having to do different stuff is what hierarchical irqdomains do best, so it sounds like a good fit. We handle some stuff at our level of the hierarchy and then fall up to the next higher level using calls such as irq_chip_ack_parent(), irq_chip_mask_parent() and friends. Yours, Linus Walleij
