Date:   Tue, 18 Sep 2018 16:41:22 +0100
From:   Marc Zyngier <marc.zyngier@....com>
To:     Guo Ren <ren_guo@...ky.com>
Cc:     <tglx@...utronix.de>, <jason@...edaemon.net>, <robh+dt@...nel.org>,
        <mark.rutland@....com>, <linux-kernel@...r.kernel.org>,
        <devicetree@...r.kernel.org>
Subject: Re: [PATCH V5 1/3] irqchip: add C-SKY irqchip drivers

On Tue, 18 Sep 2018 09:43:31 +0100,
Guo Ren <ren_guo@...ky.com> wrote:
> 
> On Mon, Sep 17, 2018 at 02:27:31PM +0100, Marc Zyngier wrote:
> > On Mon, 17 Sep 2018 03:09:29 +0100,
> > Guo Ren <ren_guo@...ky.com> wrote:
> > 
> > [...]
> > 
> > > > > +
> > > > > +	irq_set_default_host(root_domain);
> > > > 
> > > > Please drop this. There is no reason to use this on any modern, DT
> > > > based architecture.
> Ok.
> 
> > > Please let me keep this and in my arch/csky/kernel/smp.c:
> > > 
> > > void __init setup_smp_ipi(void)
> > > {
> > > 	...
> > > 	irq_create_mapping(NULL, IPI_IRQ);
> > >	rc = request_percpu_irq(IPI_IRQ, handle_ipi, "IPI Interrupt", &ipi_dummy_dev);
> > 
> > This looks quite wrong. Reading the code at
> > https://lkml.org/lkml/2018/9/12/674, it really looks like you're
> > assuming that IPI_IRQ will be mapped to a Linux IRQ with the same
> > number. Nothing could be farther from the truth.
> Yes, you are right. I should use irq_create_mapping() return value as
> the arg for request_percpu_irq. It's a stupid bug, thoug it happens to
> work.
> 
> > The Linux IRQ is returned as the result of irq_create_mapping, which
> > you're ignoring. You'd be better off creating this mapping from the
> > irqchip code, and expose the resulting Linux IRQ to oyu SMP code by
> > any mean of your choice (such as moving the send_ipi_message into the
> > irqchip code as well).
> Ok, see my diff below, is that OK?
> 
> --- a/drivers/irqchip/irq-csky-mpintc.c
> +++ b/drivers/irqchip/irq-csky-mpintc.c
> @@ -16,6 +16,7 @@
>  #include <asm/reg_ops.h>
>  #include <asm/smp.h>
>  
> +static struct irq_domain *root_domain;
>  static void __iomem *INTCG_base;
>  static void __iomem *INTCL_base;
>  
> @@ -46,7 +47,7 @@ static void csky_mpintc_handler(struct pt_regs *regs)
>  	void __iomem *reg_base = this_cpu_read(intcl_reg);
>  
>  	do {
> -		handle_domain_irq(NULL,
> +		handle_domain_irq(root_domain,
>  				  readl_relaxed(reg_base + INTCL_RDYIR),
>  				  regs);
>  	} while (readl_relaxed(reg_base + INTCL_HPPIR) & BIT(31));
> @@ -139,13 +140,17 @@ static void csky_mpintc_send_ipi(const unsigned long *mask, unsigned long irq)
>  	 */
>  	writel_relaxed((*mask) << 8 | irq, reg_base + INTCL_SIGR);
>  }
> +
> +static int csky_mpintc_ipi_irq_mapping(void)
> +{
> +	return irq_create_mapping(root_domain, IPI_IRQ);
> +}
>  #endif
>  
>  /* C-SKY multi processor interrupt controller */
>  static int __init
>  csky_mpintc_init(struct device_node *node, struct device_node *parent)
>  {
> -	struct irq_domain *root_domain;
>  	unsigned int cpu, nr_irq;
>  	int ret;
>  
> @@ -172,8 +177,6 @@ csky_mpintc_init(struct device_node *node, struct device_node *parent)
>  	if (!root_domain)
>  		return -ENXIO;
>  
> -	irq_set_default_host(root_domain);
> -
>  	/* for every cpu */
>  	for_each_present_cpu(cpu) {
>  		per_cpu(intcl_reg, cpu) = INTCL_base + (INTCL_SIZE * cpu);
> @@ -184,6 +187,8 @@ csky_mpintc_init(struct device_node *node, struct device_node *parent)
>  
>  #ifdef CONFIG_SMP
>  	set_send_ipi(&csky_mpintc_send_ipi);
> +
> +	set_ipi_irq_mapping(&csky_mpintc_ipi_irq_mapping);
>  #endif
>  
>  	return 0;
> 
> diff --git a/arch/csky/include/asm/smp.h b/arch/csky/include/asm/smp.h
> index 9a53abf..fed3a5a 100644
> --- a/arch/csky/include/asm/smp.h
> +++ b/arch/csky/include/asm/smp.h
> @@ -7,6 +7,8 @@
>  
>  #ifdef CONFIG_SMP
>  
> +#define IPI_IRQ	15
> +

It feels really bizarre that the function that maps the interrupt is
specific to the interrupt controller, and yet the interrupt number is
defined at the architecture level. I'd expect this to be just as
interrupt controller specific.

>  void __init setup_smp(void);
>  
>  void __init setup_smp_ipi(void);
> @@ -19,6 +21,8 @@ void arch_send_call_function_single_ipi(int cpu);
>  
>  void __init set_send_ipi(void (*func)(const unsigned long *, unsigned long));
>  
> +void __init set_ipi_irq_mapping(int (*func)(void));
> +
>  #define raw_smp_processor_id()	(current_thread_info()->cpu)
>  
>  #endif /* CONFIG_SMP */
> diff --git a/arch/csky/kernel/smp.c b/arch/csky/kernel/smp.c
> index 522c73f..f8343f6 100644
> --- a/arch/csky/kernel/smp.c
> +++ b/arch/csky/kernel/smp.c
> @@ -20,8 +20,6 @@
>  #include <asm/mmu_context.h>
>  #include <asm/pgalloc.h>
>  
> -#define IPI_IRQ	15
> -
>  static struct {
>  	unsigned long bits ____cacheline_aligned;
>  } ipi_data[NR_CPUS] __cacheline_aligned;
> @@ -121,13 +119,23 @@ void __init enable_smp_ipi(void)
>  	enable_percpu_irq(IPI_IRQ, 0);
>  }
>  
> +static int (*arch_ipi_irq_mapping)(void) = NULL;
> +
> +void __init set_ipi_irq_mapping(int (*func)(void))
> +{
> +	if (arch_ipi_irq_mapping)
> +		return;
> +
> +	arch_ipi_irq_mapping = func;
> +}
> +
>  void __init setup_smp_ipi(void)
>  {
> -	int rc;
> +	int rc, irq;
>  
> -	irq_create_mapping(NULL, IPI_IRQ);
> +	irq = arch_ipi_irq_mapping();

How about checking the validity of the interrupt and that
arch_ipi_irq_mapping is actually non-NULL?

>  
> -	rc = request_percpu_irq(IPI_IRQ, handle_ipi, "IPI Interrupt", &ipi_dummy_dev);
> +	rc = request_percpu_irq(irq, handle_ipi, "IPI Interrupt", &ipi_dummy_dev);
>  	if (rc)
>  		panic("%s IRQ request failed\n", __func__);

To be honest, I'd tend to question the need for this level of
abstraction, unless you actually plan for multiple SMP-capable
interrupt controllers... But at the end of the day, that's your call,
and the above code looks mostly correct.

Thanks,

	M.

-- 
Jazz is not dead, it just smell funny.

Powered by blists - more mailing lists