Re: [PATCH -tip -v6] irq_work: generic hard-irq context callbacks

[Frederic Weisbecker <fweisbec@gmail.com>]

On Tue, Sep 28, 2010 at 04:51:48PM +0800, Huang Ying wrote:
> +++ b/arch/x86/kernel/irq_work.c
> @@ -0,0 +1,30 @@
> +/*
> + * x86 specific code for irq_work
> + *
> + * Copyright (C) 2010 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
> + */
> +
> +#include <linux/kernel.h>
> +#include <linux/irq_work.h>
> +#include <linux/hardirq.h>
> +#include <asm/apic.h>
> +
> +void smp_irq_work_interrupt(struct pt_regs *regs)
> +{
> +	irq_enter();
> +	ack_APIC_irq();
> +	inc_irq_stat(apic_irq_work_irqs);
> +	irq_work_run();
> +	irq_exit();
> +}
> +
> +void arch_irq_work_raise(void)
> +{
> +#ifdef CONFIG_X86_LOCAL_APIC
> +	if (!cpu_has_apic)
> +		return;
> +
> +	apic->send_IPI_self(IRQ_WORK_VECTOR);
> +	apic_wait_icr_idle();
> +#endif
> +}



So basically, CONFIG_X86_LOCAL_APIC == !HAVE_IRQ_WORK ?

But IIUC, this will fallback to the timer interrupt:



> @@ -1279,7 +1279,10 @@ void update_process_times(int user_tick)
>  	run_local_timers();
>  	rcu_check_callbacks(cpu, user_tick);
>  	printk_tick();
> -	perf_event_do_pending();
> +#ifdef CONFIG_IRQ_WORK
> +	if (in_irq())
> +		irq_work_run();
> +#endif
>  	scheduler_tick();
>  	run_posix_cpu_timers(p);
>  }



Then HAVE_IRQ_WORK just means that the arch supports self IPIs.
So, CONFIG_IRQ_WORK doesn't need to depend on HAVE_IRQ_WORK
because of the timer fallback. But archs that support self
IPIs should avoid the above fallback because it bloats
the timer interrupt.

Perhaps CONFIG_HAVE_IRQ_WORK should be CONFIG_HAVE_IRQ_WORK_SOURCE
or CONFIG_HAVE_IRQ_WORK_TRIGGER to better denote the capability.
And then:

config IRQ_WORK_TRIGGER
	depends on HAVE_IRQ_WORK_TRIGGER && IRQ_WORK
	default y

And that would define the right condition to build the fallback
in the timer interrupt.

Or I am completely misunderstanding something?

Thanks.