Re: [PATCH v5 3/7] genirq: Add mechanism to multiplex a single HW IPI

[Anup Patel]

On Mon, Apr 11, 2022 at 1:41 AM Thomas Gleixner <tglx@xxxxxxxxxxxxx> wrote:
>
> Anup,
>
> On Thu, Mar 24 2022 at 20:42, 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 need a generic mechanism to
> > create multiple per-CPU vIRQs using a single HW IPI hence this patch.
>
> git grep 'This patch' Documentation/process

Okay, I will update the commit description as-per Documentation.

>
> > The generic IPI multiplex mechanism added by this patch can also be
> > useful to other architectures.
>
> Which ones? Sane architectures have more than one IPI.

Currently, the IPI muxing is shared code for various RISC-V drivers
(such as CLINT driver, SBI IPI irqchip driver, and AIA (coming soon)).

Overall, the IPI muxing seems independent of RISC-V so maybe
it is useful to have it as common selectable API.

>
> > diff --git a/include/linux/irq.h b/include/linux/irq.h
> > index 848e1e12c5c6..cdce7eae2f37 100644
> > --- a/include/linux/irq.h
> > +++ b/include/linux/irq.h
> > @@ -1248,6 +1248,34 @@ int __ipi_send_mask(struct irq_desc *desc, const struct cpumask *dest);
> >  int ipi_send_single(unsigned int virq, unsigned int cpu);
> >  int ipi_send_mask(unsigned int virq, const struct cpumask *dest);
> >
> > +#define IPI_MUX_NR_IRQS              BITS_PER_LONG
> > +struct ipi_mux_ops {
>
> This is unreadable. Newlines exist for a reason.

Okay, I will add a newline above the struct.

>
> > +     void (*ipi_mux_clear)(unsigned int parent_virq);
> > +     void (*ipi_mux_send)(unsigned int parent_virq,
> > +                          const struct cpumask *mask);
> > +};
> > +
> > +/* Process multiplexed IPIs */
> > +void ipi_mux_process(void);
> > +
> > +/*
> > + * Create multiple IPIs (total IPI_MUX_NR_IRQS) multiplexed on top of a
> > + * single parent IPI.
> > + *
> > + * If the parent IPI > 0 then ipi_mux_process() will be automatically
> > + * called via chained handler.
> > + *
> > + * If the parent IPI <= 0 then it is responsiblity of irqchip drivers
> > + * to explicitly call ipi_mux_process() for processing muxed
> > + * IPIs.
> > + *
> > + * Returns first virq of the muxed IPIs upon success or <=0 upon failure
> > + */
> > +int ipi_mux_create(unsigned int parent_virq, const struct ipi_mux_ops *ops);
>
> While it is kinda sensible to have the documentation near the
> declaration, I prefer it to be near the code because thats where it
> matters and also has a higher chance to be updated when the code
> changes.

Okay, I will move documentation near the code.

>
> Please use proper kernel doc while at it.

Sure, I will update.

>
> > +static unsigned int ipi_mux_parent_virq;
> > +static struct irq_domain *ipi_mux_domain;
> > +static const struct  ipi_mux_ops *ipi_mux_ops;
> > +static DEFINE_PER_CPU(unsigned long, ipi_mux_bits);
> > +
> > +static void ipi_mux_dummy(struct irq_data *d)
> > +{
> > +}
> > +
> > +static void ipi_mux_send_mask(struct irq_data *d, const struct cpumask *mask)
> > +{
> > +     int cpu;
> > +
> > +     /* Barrier before doing atomic bit update to IPI bits */
> > +     smp_mb__before_atomic();
> > +
> > +     for_each_cpu(cpu, mask)
> > +             set_bit(d->hwirq, per_cpu_ptr(&ipi_mux_bits, cpu));
> > +
> > +     /* Barrier after doing atomic bit update to IPI bits */
> > +     smp_mb__after_atomic();
> > +
> > +     /* Trigger the parent IPI */
> > +     ipi_mux_ops->ipi_mux_send(ipi_mux_parent_virq, mask);
> > +}
> > +
> > +static struct irq_chip ipi_mux_chip = {
> > +     .name           = "RISC-V IPI Mux",
>
> RISC-V IPI Mux is a truly generic name :)

Aargh, I forgot to remove "RISC-V" from the name here. I will update.

>
> > +     .irq_mask       = ipi_mux_dummy,
> > +     .irq_unmask     = ipi_mux_dummy,
> > +     .ipi_send_mask  = ipi_mux_send_mask,
> > +};
> > +
> > +static int ipi_mux_domain_map(struct irq_domain *d, unsigned int irq,
> > +                           irq_hw_number_t hwirq)
> > +{
> > +     irq_set_percpu_devid(irq);
> > +     irq_domain_set_info(d, irq, hwirq, &ipi_mux_chip, d->host_data,
> > +                         handle_percpu_devid_irq, NULL, NULL);
> > +
> > +     return 0;
> > +}
> > +
> > +static int ipi_mux_domain_alloc(struct irq_domain *d, unsigned int virq,
> > +                             unsigned int nr_irqs, void *arg)
> > +{
> > +     int i, ret;
> > +     irq_hw_number_t hwirq;
> > +     unsigned int type = IRQ_TYPE_NONE;
> > +     struct irq_fwspec *fwspec = arg;
>
>   Documentation/process/maintainer-tip.rst #coding-style-notes

Okay, I will refer and update.

>
> > +     ret = irq_domain_translate_onecell(d, fwspec, &hwirq, &type);
> > +     if (ret)
> > +             return ret;
> > +
> > +     for (i = 0; i < nr_irqs; i++) {
> > +             ret = ipi_mux_domain_map(d, virq + i, hwirq + i);
> > +             if (ret)
> > +                     return ret;
> > +     }
> > +
> > +     return 0;
> > +}
> > +
> > +static const struct irq_domain_ops ipi_mux_domain_ops = {
> > +     .translate      = irq_domain_translate_onecell,
> > +     .alloc          = ipi_mux_domain_alloc,
> > +     .free           = irq_domain_free_irqs_top,
> > +};
> > +
> > +void ipi_mux_process(void)
> > +{
> > +     int err;
> > +     unsigned long irqs, *bits = this_cpu_ptr(&ipi_mux_bits);
> > +     irq_hw_number_t hwirq;
> > +
> > +     while (true) {
> > +             /* Clear the parent IPI */
> > +             ipi_mux_ops->ipi_mux_clear(ipi_mux_parent_virq);
>
> This being in a loop smells fishy at least without a comment. And the
> more I read all of this the less I'm convinced that this code can be
> used by anything else than RISCV.

The original IPI muxing code in RISC-V had this loop so I did not
remove it.

Actually, the loop is redundant because if a CPU gets another IPI
while it was in ipi_mux_process() then another interrupt will be taken
and ipi_mux_process() will be called again. I test more and remove
this loop.

>
> > +             /* Order bit clearing and data access. */
> > +             mb();
>
> This mb() pairs with what? Memory barriers have a counterpart and it's
> mandatory to document that in the comment.

It pairs with barriers in ipi_mux_send_mask(). I will update the comment.

>
> > +             irqs = xchg(bits, 0);
> > +             if (!irqs)
> > +                     break;
> > +
> > +             for_each_set_bit(hwirq, &irqs, IPI_MUX_NR_IRQS) {
> > +                     err = generic_handle_domain_irq(ipi_mux_domain,
> > +                                                     hwirq);
> > +                     if (unlikely(err))
> > +                             pr_warn_ratelimited(
> > +                                     "can't find mapping for hwirq %lu\n",
> > +                                     hwirq);
> > +             }
> > +     }
> > +}
> > +
> > +
> > +void ipi_mux_destroy(void)
>
> Seriously? You provide a function to rip the IPI mechanism out in a
> running system? What's that for?
>
> > +{
> > +     if (!ipi_mux_domain)
> > +             return;
> > +
> > +     irq_domain_remove(ipi_mux_domain);
> > +     ipi_mux_domain = NULL;
> > +     ipi_mux_parent_virq = 0;
>
> If it would be useful, then this would leak the hotplug callbacks, but
> the good news is that after tearing down the IPI domain hotplug does not
> work anymore :)

The only use of this function was to clean up in-case the irqchip
driver failed after creating mux.

I will certainly remove this function in the next patch revision.

>
> Thanks,
>
>         tglx

Regards,
Anup