Re: [RFC][PATCH] irq_work

[Huang Ying <ying.huang@intel.com>]

On Thu, 2010-06-24 at 14:35 +0800, Peter Zijlstra wrote:
> Something like this, but filled out with some arch code that does the
> self-ipi and calls irq_work_run() should do.
> 
> No need to molest the softirq code, no need for limited vectors of any
> kind.

Now, as far as my understanding goes, hard IRQ based solution is
acceptable for everyone.

Ingo and Andi,

Do you agree?

> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
> ---
>  include/linux/irq_callback.h |   13 ++++++++
>  kernel/irq_callback.c        |   66 +++++++++++++++++++++++++++++++++++++++++++
>  2 files changed, 79 insertions(+)
> 
> Index: linux-2.6/include/linux/irq_callback.h
> ===================================================================
> --- /dev/null
> +++ linux-2.6/include/linux/irq_callback.h
> @@ -0,0 +1,13 @@
> +#ifndef _LINUX_IRQ_CALLBACK_H
> +#define _LINUX_IRQ_CALLBACK_H
> +
> +struct irq_work {
> +	struct irq_work *next;
> +	void (*func)(struct irq_work *);
> +};

It is better to add "void *data" field in this struct to allow same
function can be used for multiple struct irq_work.

And I think IRQ is the implementation detail here, so irq_work is
probably not a good name. nmi_return_notifier or nmi_callback is better?

> +int irq_work_queue(struct irq_work *entry, void (*func)(struct irq_work *));
> +void irq_work_run(void);
> +void irq_work_sync(struct irq_work *entry);
> +
> +#endif /* _LINUX_IRQ_CALLBACK_H */
> Index: linux-2.6/kernel/irq_callback.c
> ===================================================================
> --- /dev/null
> +++ linux-2.6/kernel/irq_callback.c
> @@ -0,0 +1,66 @@
> +
> +#include <linux/irq_callback.h>
> +
> +#define CALLBACK_TAIL ((struct irq_work *)-1UL)
> +
> +static DEFINE_PER_CPU(struct irq_work *, irq_work_list) = {
> +	CALLBACK_TAIL,
> +};
> +
> +int irq_work_queue(struct irq_work *entry, void (*func)(struct irq_work *))
> +{
> +	struct irq_work **head;
> +
> +	if (cmpxchg(&entry->next, NULL, CALLBACK_TAIL) != NULL)
> +		return 0;
> +
> +	entry->func = func;
> +
> +	head = &get_cpu_var(irq_work_list);
> +
> +	do {
> +		entry->next = *head;
> +	} while (cmpxchg(head, entry->next, entry) != entry->next);
> +
> +	if (entry->next == CALLBACK_TAIL)
> +		arch_self_ipi();
> +
> +	put_cpu_var(irq_work_list);
> +	return 1;
> +}
> +
> +void irq_work_run(void)
> +{
> +	struct irq_work *list;
> +
> +	list = xchg(&__get_cpu_var(irq_work_list), CALLBACK_TAIL);
> +	while (list != CALLBACK_TAIL) {
> +		struct irq_work *entry = list;
> +
> +		list = list->next;
> +		entry->func(entry);
> +
> +		entry->next = NULL;

entry->next = NULL should be put before entry->func(entry), so that we
will not lose a notification from NMI. And maybe check irq_work_list for
several times to make sure nothing is lost.

> +		/*
> +		 * matches the mb in cmpxchg() in irq_work_queue()
> +		 */
> +		smp_wmb();
> +	}
> +}

I don't know why we need smp_wmb() here and smp_rmb() in
irq_work_pending(). The smp_<x>mb() in original perf_pending_xxx code is
not necessary too. Because smp_<x>mb is invoked in wake_up_process() and
__wait_event() already.

> +static int irq_work_pending(struct irq_work *entry)
> +{
> +	/*
> +	 * matches the wmb in irq_work_run
> +	 */
> +	smp_rmb();
> +	return entry->next != NULL;
> +}
> +
> +void irq_work_sync(struct irq_work *entry)
> +{
> +	WARN_ON_ONCE(irqs_disabled());
> +
> +	while (irq_work_pending(entry))
> +		cpu_relax();
> +}

If we move entry->next = NULL earlier in irq_work_run(), we need another
flag to signify the entry->func is running here.

Best Regards,
Huang Ying