Re: [PATCH 05/13] irqchip: add initial support for ompic

[Marc Zyngier]

On 30/08/17 22:58, Stafford Horne wrote:
> From: Stefan Kristiansson <stefan.kristiansson@xxxxxxxxxxxxx>
> 
> IPI driver for OpenRISC Multicore programmable interrupt controller as
> described in the Multicore support section of the OpenRISC 1.2
> proposed architecture specification:
> 
>   https://github.com/stffrdhrn/doc/raw/arch-1.2-proposal/openrisc-arch-1.2-rev0.pdf
> 
> Each OpenRISC core contains a full interrupt controller which is used in
> the SMP architecture for interrupt balancing.  This IPI device is the
> only external device required for enabling SMP on OpenRISC.
> 
> Pending ops are stored in a memory bit mask which can allow multiple
> pending operations to be set and serviced at a time. This is mostly
> borrowed from the alpha IPI implementation.
> 
> Signed-off-by: Stefan Kristiansson <stefan.kristiansson@xxxxxxxxxxxxx>
> [shorne@xxxxxxxxx: converted ops to bitmask, wrote commit message]
> Signed-off-by: Stafford Horne <shorne@xxxxxxxxx>
> ---
>  .../bindings/interrupt-controller/ompic.txt        |  22 ++++
>  arch/openrisc/Kconfig                              |   1 +
>  drivers/irqchip/Kconfig                            |   4 +
>  drivers/irqchip/Makefile                           |   1 +
>  drivers/irqchip/irq-ompic.c                        | 117 +++++++++++++++++++++
>  5 files changed, 145 insertions(+)
>  create mode 100644 Documentation/devicetree/bindings/interrupt-controller/ompic.txt
>  create mode 100644 drivers/irqchip/irq-ompic.c
> 
> diff --git a/Documentation/devicetree/bindings/interrupt-controller/ompic.txt b/Documentation/devicetree/bindings/interrupt-controller/ompic.txt
> new file mode 100644
> index 000000000000..4176ecc3366d
> --- /dev/null
> +++ b/Documentation/devicetree/bindings/interrupt-controller/ompic.txt
> @@ -0,0 +1,22 @@
> +OpenRISC Multicore Programmable Interrupt Controller
> +
> +Required properties:
> +
> +- compatible : This should be "ompic"
> +- reg : Specifies base physical address and size of the register space. The
> +  size can be arbitrary based on the number of cores the controller has
> +  been configured to handle, typically 8 bytes per core.
> +- interrupt-controller : Identifies the node as an interrupt controller
> +- #interrupt-cells : Specifies the number of cells needed to encode an
> +  interrupt source. The value shall be 1.
> +- interrupts : Specifies the interrupt line to which the ompic is wired.
> +
> +Example:
> +
> +ompic: ompic {
> +	compatible = "ompic";
> +	reg = <0x98000000 16>;
> +	#interrupt-cells = <1>;
> +	interrupt-controller;
> +	interrupts = <1>;
> +};
> diff --git a/arch/openrisc/Kconfig b/arch/openrisc/Kconfig
> index 214c837ce597..dd7e55e7e42d 100644
> --- a/arch/openrisc/Kconfig
> +++ b/arch/openrisc/Kconfig
> @@ -30,6 +30,7 @@ config OPENRISC
>  	select NO_BOOTMEM
>  	select ARCH_USE_QUEUED_SPINLOCKS
>  	select ARCH_USE_QUEUED_RWLOCKS
> +	select OMPIC if SMP
>  
>  config CPU_BIG_ENDIAN
>  	def_bool y
> diff --git a/drivers/irqchip/Kconfig b/drivers/irqchip/Kconfig
> index f1fd5f44d1d4..3fa60e6667a7 100644
> --- a/drivers/irqchip/Kconfig
> +++ b/drivers/irqchip/Kconfig
> @@ -145,6 +145,10 @@ config CLPS711X_IRQCHIP
>  	select SPARSE_IRQ
>  	default y
>  
> +config OMPIC
> +	bool
> +	select IRQ_DOMAIN

Why do you need to select IRQ_DOMAIN? This driver doesn't seem to use any...

> +
>  config OR1K_PIC
>  	bool
>  	select IRQ_DOMAIN
> diff --git a/drivers/irqchip/Makefile b/drivers/irqchip/Makefile
> index e88d856cc09c..123047d7a20d 100644
> --- a/drivers/irqchip/Makefile
> +++ b/drivers/irqchip/Makefile
> @@ -17,6 +17,7 @@ obj-$(CONFIG_DW_APB_ICTL)		+= irq-dw-apb-ictl.o
>  obj-$(CONFIG_METAG)			+= irq-metag-ext.o
>  obj-$(CONFIG_METAG_PERFCOUNTER_IRQS)	+= irq-metag.o
>  obj-$(CONFIG_CLPS711X_IRQCHIP)		+= irq-clps711x.o
> +obj-$(CONFIG_OMPIC)			+= irq-ompic.o
>  obj-$(CONFIG_OR1K_PIC)			+= irq-or1k-pic.o
>  obj-$(CONFIG_ORION_IRQCHIP)		+= irq-orion.o
>  obj-$(CONFIG_OMAP_IRQCHIP)		+= irq-omap-intc.o
> diff --git a/drivers/irqchip/irq-ompic.c b/drivers/irqchip/irq-ompic.c
> new file mode 100644
> index 000000000000..438819f8a5a7
> --- /dev/null
> +++ b/drivers/irqchip/irq-ompic.c
> @@ -0,0 +1,117 @@
> +/*
> + * Open Multi-Processor Interrupt Controller driver
> + *
> + * Copyright (C) 2014 Stefan Kristiansson <stefan.kristiansson@xxxxxxxxxxxxx>
> + *
> + * This file is licensed under the terms of the GNU General Public License
> + * version 2.  This program is licensed "as is" without any warranty of any
> + * kind, whether express or implied.
> + */
> +
> +#include <linux/io.h>
> +#include <linux/interrupt.h>
> +#include <linux/smp.h>
> +#include <linux/of.h>
> +#include <linux/of_irq.h>
> +#include <linux/of_address.h>
> +#include <linux/irqchip/chained_irq.h>

Don't think you need this.

> +#include <linux/delay.h>

Nor this.

> +
> +#include <linux/irqchip.h>
> +
> +#define OMPIC_IPI_BASE			0x0
> +#define OMPIC_IPI_CTRL(cpu)		(OMPIC_IPI_BASE + 0x0 + (cpu)*8)
> +#define OMPIC_IPI_STAT(cpu)		(OMPIC_IPI_BASE + 0x4 + (cpu)*8)

In the DT binding you say that "size can be arbitrary based on the
number of cores the controller has been configured to handle, typically
8 bytes per core". Here, this is 8 bytes, always, which is not exactly
the same. What is the architectural value, if any? If there is none,
then the per-core size should either come from DT or some other mean
(register?).

> +
> +#define OMPIC_IPI_CTRL_IRQ_ACK		(1 << 31)
> +#define OMPIC_IPI_CTRL_IRQ_GEN		(1 << 30)
> +#define OMPIC_IPI_CTRL_DST(cpu)		(((cpu) & 0x3fff) << 16)
> +
> +#define OMPIC_IPI_STAT_IRQ_PENDING	(1 << 30)
> +
> +#define OMPIC_IPI_DATA(x)		((x) & 0xffff)
> +
> +static struct {
> +	unsigned long ops;
> +} ipi_data[NR_CPUS];

In general, a per cpu data structure is better expressed as a percpu
data structure (yes, I'm in a funny mood this morning). Otherwise,
you're going to thrash more than just the receiver and the sender, but
also all the other CPUs that have their data in the same cache line.

> +
> +static void __iomem *ompic_base;
> +
> +static inline u32 ompic_readreg(void __iomem *base, loff_t offset)
> +{
> +	return ioread32be(base + offset);
> +}
> +
> +static void ompic_writereg(void __iomem *base, loff_t offset, u32 data)
> +{
> +	iowrite32be(data, base + offset);
> +}
> +
> +#ifdef CONFIG_SMP

This code is only selected when CONFIG_SMP=y.

> +void ompic_raise_softirq(const struct cpumask *mask, unsigned int irq)
> +{

What is "irq" here? How is it guaranteed to fit in an unsigned long?

> +	unsigned int dst_cpu;
> +	unsigned int src_cpu = smp_processor_id();
> +
> +	for_each_cpu(dst_cpu, mask) {
> +		set_bit(irq, &ipi_data[dst_cpu].ops);
> +
> +		ompic_writereg(ompic_base, OMPIC_IPI_CTRL(src_cpu),
> +			       OMPIC_IPI_CTRL_IRQ_GEN |
> +			       OMPIC_IPI_CTRL_DST(dst_cpu) |
> +			       OMPIC_IPI_DATA(1));

What guarantees that the action of set_bit can be observed by the target
CPU? set-bit gives you atomicity, but no barrier.

> +	}
> +}
> +#endif
> +
> +irqreturn_t ompic_ipi_handler(int irq, void *dev_id)
> +{
> +	unsigned int cpu = smp_processor_id();
> +	unsigned long *pending_ops = &ipi_data[cpu].ops;
> +	unsigned long ops;
> +
> +	ompic_writereg(ompic_base, OMPIC_IPI_CTRL(cpu), OMPIC_IPI_CTRL_IRQ_ACK);
> +	while ((ops = xchg(pending_ops, 0)) != 0) {

Barrier again?

> +		do {
> +			unsigned long ipi;
> +
> +			ipi = ops & -ops;
> +			ops &= ~ipi;
> +			ipi = __ffs(ipi);

This feels pointlessly convoluted. Is there any reason why you can't
write it as:

			ipi = __ffs(ops);
			ops &= ~(1UL << ipi);

which feels like a much more common idiom?

> +
> +			handle_IPI(ipi);
> +		} while (ops);
> +	}
> +
> +	return IRQ_HANDLED;
> +}
> +
> +static struct irqaction ompi_ipi_irqaction = {
> +	.handler =      ompic_ipi_handler,
> +	.flags =        IRQF_PERCPU,
> +	.name =         "ompic_ipi",
> +};
> +
> +#ifdef CONFIG_OF

This code is useless if you don't have CONFIG_OF. So either you make the
driver depend on CONFIG_OF, or you actively select it. And given that
CONFIG_OPENRISC already selects CONFIG_OF, you can happily get rid of
all the #ifdefery here.

> +int __init ompic_of_init(struct device_node *node, struct device_node *parent)
> +{
> +	int irq;
> +
> +	if (WARN_ON(!node))
> +		return -ENODEV;

How do you end-up here if node == NULL?

> +
> +	memset(ipi_data, 0, sizeof(ipi_data));

The kernel should do this for you already.

> +
> +	ompic_base = of_iomap(node, 0);
> +
> +	irq = irq_of_parse_and_map(node, 0);
> +	setup_irq(irq, &ompi_ipi_irqaction);
> +
> +#ifdef CONFIG_SMP
> +	set_smp_cross_call(ompic_raise_softirq);
> +#endif
> +
> +	return 0;
> +}
> +IRQCHIP_DECLARE(ompic, "ompic", ompic_of_init);
> +#endif
> 

Thanks,

	M.
-- 
Jazz is not dead. It just smells funny...