On Wed, Oct 25 2017 at 5:37:24 pm BST, Paul Burton <paul.burton@xxxxxxxx> wrote:
> The gic_all_vpes_local_irq_controller chip currently attempts to operate
> on all CPUs/VPs in the system when masking or unmasking an interrupt.
> This has a few drawbacks:
>
> - In multi-cluster systems we may not always have access to all CPUs in
> the system. When all CPUs in a cluster are powered down that
> cluster's GIC may also power down, in which case we cannot configure
> its state.
>
> - Relatedly, if we power down a cluster after having configured
> interrupts for CPUs within it then the cluster's GIC may lose state &
> we need to reconfigure it. The current approach doesn't take this
> into account.
>
> - It's wasteful if we run Linux on fewer VPs than are present in the
> system. For example if we run a uniprocessor kernel on CPU0 of a
> system with 16 CPUs then there's no point in us configuring CPUs
> 1-15.
>
> - The implementation is also lacking in that it expects the range
> 0..gic_vpes-1 to represent valid Linux CPU numbers which may not
> always be the case - for example if we run on a system with more VPs
> than the kernel is configured to support.
>
> Fix all of these issues by only configuring the affected interrupts for
> CPUs which are online at the time, and recording the configuration in a
> new struct gic_all_vpes_chip_data for later use by CPUs being brought
> online. We register a CPU hotplug state (reusing
> CPUHP_AP_IRQ_GIC_STARTING which the ARM GIC driver uses, and which seems
> suitably generic for reuse with the MIPS GIC) and execute
> irq_cpu_online() in order to configure the interrupts on the newly
> onlined CPU.
>
> Signed-off-by: Paul Burton <paul.burton@xxxxxxxx>
> Cc: Jason Cooper <jason@xxxxxxxxxxxxxx>
> Cc: Marc Zyngier <marc.zyngier@xxxxxxx>
> Cc: Thomas Gleixner <tglx@xxxxxxxxxxxxx>
> Cc: linux-mips@xxxxxxxxxxxxxx
> ---
>
> drivers/irqchip/irq-mips-gic.c | 72 ++++++++++++++++++++++++++++++++----------
> 1 file changed, 56 insertions(+), 16 deletions(-)
>
> diff --git a/drivers/irqchip/irq-mips-gic.c b/drivers/irqchip/irq-mips-gic.c
> index 6fdcc1552fab..dd9da773db90 100644
> --- a/drivers/irqchip/irq-mips-gic.c
> +++ b/drivers/irqchip/irq-mips-gic.c
[...]
> @@ -622,6 +653,13 @@ static const struct irq_domain_ops gic_ipi_domain_ops = {
> .match = gic_ipi_domain_match,
> };
>
> +static int gic_cpu_startup(unsigned int cpu)
> +{
> + /* Invoke irq_cpu_online callbacks to enable desired interrupts */
> + irq_cpu_online();
> +
> + return 0;
> +}
>
> static int __init gic_of_init(struct device_node *node,
> struct device_node *parent)
> @@ -768,6 +806,8 @@ static int __init gic_of_init(struct device_node *node,
> }
> }
>
> - return 0;
> + return cpuhp_setup_state(CPUHP_AP_IRQ_GIC_STARTING,
> + "irqchip/mips/gic:starting",
> + gic_cpu_startup, NULL);
I'm wondering about this. CPUHP_AP_IRQ_GIC_STARTING is a symbol that is
used on ARM platforms. You're very welcome to use it (as long as nobody
builds a system with both an ARM GIC and a MIPS GIC...), but I'm a bit
worried that we could end-up breaking things if one of us decides to
reorder it in enum cpuhp_state.
The safest option would be for you to add your own state value, which
would allow the two architecture to evolve independently.
Thoughts?
M.
--
Jazz is not dead. It just smells funny.
