Re: [RFC PATCH] timekeeping: introduce timekeeping_is_busy()

[Peter Zijlstra <peterz@infradead.org>]

On Wed, Sep 11, 2013 at 08:48:11PM -0400, Mathieu Desnoyers wrote:
> Thoughts ?


struct foo {
	...
};

spinlock_t foo_lock;
unsigned int foo_head = 0;
unsigned int foo_tail = 0;
struct foo foo_array[2];

void foo_assign(struct foo f)
{
	spin_lock(&foo_lock);
	foo_head++;
	smp_wmb();
	foo_array[foo_head & 1] = f;
	smp_wmb();
	foo_tail++;
	spin_unlock(&foo_lock);
}

struct foo foo_get(void)
{
	unsigned int tail, head;
	struct foo ret;

again:
	tail = ACCESS_ONCE(foo_tail);
	smp_rmb();
	ret = foo_array[tail & 1];
	smp_rmb();
	head = ACCESS_ONCE(foo_head);
	if (head - tail >= 2)
		goto again;

	return ret;
}

Should work and get you the most recent 'complete' foo even when
foo_get() is called nested inside foo_assign().