Re: [PATCH v14 3/8] genirq: Add mechanism to multiplex a single HW IPI
From: Thomas Gleixner
Date: Thu Dec 01 2022 - 12:20:37 EST
On Thu, Dec 01 2022 at 18:31, Anup Patel wrote:
> All RISC-V platforms have a single HW IPI provided by the INTC local
> interrupt controller. The HW method to trigger INTC IPI can be through
> external irqchip (e.g. RISC-V AIA), through platform specific device
> (e.g. SiFive CLINT timer), or through firmware (e.g. SBI IPI call).
>
> To support multiple IPIs on RISC-V, we add a generic IPI multiplexing
s/we//
> mechanism which help us create multiple virtual IPIs using a single
> HW IPI. This generic IPI multiplexing is inspired from the Apple AIC
s/from/by/
> irqchip driver and it is shared by various RISC-V irqchip drivers.
Sure, but now we have two copies of this. One in the Apple AIC and one
here. The obvious thing to do is:
1) Provide generic infrastructure
2) Convert AIC to use it
3) Add RISCV users
No?
> +static void ipi_mux_mask(struct irq_data *d)
> +{
> + struct ipi_mux_cpu *icpu = this_cpu_ptr(ipi_mux_pcpu);
> +
> + atomic_andnot(BIT(irqd_to_hwirq(d)), &icpu->enable);
> +}
> +
> +static void ipi_mux_unmask(struct irq_data *d)
> +{
> + u32 ibit = BIT(irqd_to_hwirq(d));
> + struct ipi_mux_cpu *icpu = this_cpu_ptr(ipi_mux_pcpu);
The AIC code got the variable ordering correct ...
https://www.kernel.org/doc/html/latest/process/maintainer-tip.html#variable-declarations
> + atomic_or(ibit, &icpu->enable);
> +
> + /*
> + * The atomic_or() above must complete before the atomic_read()
> + * below to avoid racing ipi_mux_send_mask().
> + */
> + smp_mb__after_atomic();
> +
> + /* If a pending IPI was unmasked, raise a parent IPI immediately. */
> + if (atomic_read(&icpu->bits) & ibit)
> + ipi_mux_send(smp_processor_id());
> +}
> +
> +static void ipi_mux_send_mask(struct irq_data *d, const struct cpumask *mask)
> +{
> + u32 ibit = BIT(irqd_to_hwirq(d));
> + struct ipi_mux_cpu *icpu = this_cpu_ptr(ipi_mux_pcpu);
> + unsigned long pending;
> + int cpu;
> +
> + for_each_cpu(cpu, mask) {
> + icpu = per_cpu_ptr(ipi_mux_pcpu, cpu);
> + pending = atomic_fetch_or_release(ibit, &icpu->bits);
> +
> + /*
> + * The atomic_fetch_or_release() above must complete
> + * before the atomic_read() below to avoid racing with
> + * ipi_mux_unmask().
> + */
> + smp_mb__after_atomic();
> +
> + /*
> + * The flag writes must complete before the physical IPI is
> + * issued to another CPU. This is implied by the control
> + * dependency on the result of atomic_read() below, which is
> + * itself already ordered after the vIPI flag write.
> + */
> + if (!(pending & ibit) && (atomic_read(&icpu->enable) & ibit))
> + ipi_mux_send(cpu);
> + }
> +}
> +
> +static const struct irq_chip ipi_mux_chip = {
> + .name = "IPI Mux",
> + .irq_mask = ipi_mux_mask,
> + .irq_unmask = ipi_mux_unmask,
> + .ipi_send_mask = ipi_mux_send_mask,
> +};
> +
> +static int ipi_mux_domain_alloc(struct irq_domain *d, unsigned int virq,
> + unsigned int nr_irqs, void *arg)
> +{
> + int i;
> +
> + for (i = 0; i < nr_irqs; i++) {
> + irq_set_percpu_devid(virq + i);
> + irq_domain_set_info(d, virq + i, i, &ipi_mux_chip, NULL,
> + handle_percpu_devid_irq, NULL, NULL);
> + }
> +
> + return 0;
> +}
> +
> +static const struct irq_domain_ops ipi_mux_domain_ops = {
> + .alloc = ipi_mux_domain_alloc,
> + .free = irq_domain_free_irqs_top,
> +};
> +
> +/**
> + * ipi_mux_process - Process multiplexed virtual IPIs
> + */
> +void ipi_mux_process(void)
> +{
> + struct ipi_mux_cpu *icpu = this_cpu_ptr(ipi_mux_pcpu);
> + irq_hw_number_t hwirq;
> + unsigned long ipis;
> + unsigned int en;
> +
> + /*
> + * Reading enable mask does not need to be ordered as long as
> + * this function called from interrupt handler because only
> + * the CPU itself can change it's own enable mask.
> + */
> + en = atomic_read(&icpu->enable);
> +
> + /*
> + * Clear the IPIs we are about to handle. This pairs with the
> + * atomic_fetch_or_release() in ipi_mux_send_mask().
The comments in the AIC code where you copied from are definitely
better...
Thanks,
tglx