Locking in the clk API, part 2: clk_prepare/clk_unprepare

On Tue, Feb 01, 2011 at 11:54:49AM +0100, Uwe Kleine-K?nig wrote:

Alternatively don't force the sleep in clk_prepare (e.g. by protecting
prepare_count by a spinlock (probably enable_lock)) and call clk_prepare
before calling clk->ops->enable?

That's a completely bad idea.  I assume you haven't thought about this
very much.

There's two ways I can think of doing what you're suggesting:

int clk_prepare(struct clk *clk)
{
	unsigned long flags;
	int ret = 0;

	might_sleep();

	spin_lock_irqsave(&clk->enable_lock, flags);
	if (clk->prepare_count++ == 0)
		ret = clk->ops->prepare(clk);
	spin_unlock_irqrestore(&clk->enable_clock, flags);

	return ret;
}

The problem is that clk->ops->prepare() is called in a non-sleepable
context.  So this breaks the whole idea of clk_prepare(), and so isn't
a solution.

The other solution is:

int clk_prepare(struct clk *clk)
{
	unsigned long flags;
	int ret = 0;
	bool first;

	might_sleep();

	spin_lock_irqsave(clk->enable_lock, flags);
	first = clk->prepare_count++ == 0;
	spin_unlock_irqrestore(clk->enable_clock, flags);

	if (first)
		ret = clk->ops->prepare(clk);

	return ret;
}

The problem with this is that you now don't have any sane locking on
the prepare callback, and the circumstances under which it's called
are very indefinite.  For example, consider a preempt-enabled system:

	thread 1		thread 2		prepare_count
	clk_prepare					0
	clk->prepare_count++				1
	<thread switch>
				clk_prepare		1
				clk->prepare_count++	2
				clk_prepare returns	2

				clk_enable		2
				<explodes as clock is not prepared>
				<thread switch>
	clk->ops->prepare(clk)

So really, what you're suggesting is completely broken.