Re: [PATCH v4 07/15] lockdep: Implement crossrelease feature

[Peter Zijlstra <peterz@infradead.org>]

On Tue, Jan 17, 2017 at 11:05:42AM +0900, Byungchul Park wrote:
> On Mon, Jan 16, 2017 at 04:10:01PM +0100, Peter Zijlstra wrote:
> > On Fri, Dec 09, 2016 at 02:12:03PM +0900, Byungchul Park wrote:

> > > +
> > > +	/*
> > > +	 * Whenever irq happens, these are updated so that we can
> > > +	 * distinguish each irq context uniquely.
> > > +	 */
> > > +	unsigned int		hardirq_id;
> > > +	unsigned int		softirq_id;
> > 
> > An alternative approach would be to 'unwind' or discard all historical
> > events from a nested context once we exit it.
> 
> That's one of what I considered. However, it would make code complex to
> detect if pend_lock ring buffer was wrapped.

I'm not sure I see the need for detecting that...

> > 
> > After all, all we care about is the history of the release context, once
> > the context is gone, we don't care.
> 
> We must care it and decide if the next plock in the ring buffer might be
> valid one or not.

So I was thinking this was an overwriting ring buffer; something like
so:

struct pend_lock plocks[64];
unsigned int plocks_idx;

static void plocks_add(..)
{
	unsigned int idx = (plocks_idx++) % 64;

	plocks[idx] = ...;
}

static void plocks_close_context(int ctx)
{
	for (i = 0; i < 64; i++) {
		int idx = (plocks_idx - 1) % 64;
		if (plocks[idx].ctx != ctx)
			break;

		plocks_idx--;
	}
}

Similarly for the release, it need only look at 64 entries and terminate
early if the generation number is too old.

static void plocks_release(unsigned int gen)
{
	for (i = 0; i < 64; i++) {
		int idx = (plocks_idx - 1 - i) % 64;
		if ((int)(plocks[idx].gen_id - gen) < 0)
			break;

		/* do release muck */
	}
}