Re: [RFC PATCH] locking/percpu-rwsem: use this_cpu_{inc|dec}() for read_count

[Oleg Nesterov]

On 09/18, Peter Zijlstra wrote:
>
> On Fri, Sep 18, 2020 at 12:01:12PM +0200, peterz@xxxxxxxxxxxxx wrote:
> > +	u64 sum = per_cpu_sum(*(u64 *)sem->read_count);
>
> Moo, that doesn't work, we have to do two separate sums.

Or we can re-introduce "atomic_t slow_read_ctr".

	percpu_up_read_irqsafe(sem)
	{
		preempt_disable();
		atomic_dec_release(&sem->slow_read_ctr);
		if (!rcu_sync_is_idle(&sem->rss))
			rcuwait_wake_up(&sem->writer);
		preempt_enable();
	}

	readers_active_check(sem)
	{
		unsigned int sum = per_cpu_sum(*sem->read_count) +
			(unsigned int)atomic_read(&sem->slow_read_ctr);
		if (sum)
			return false;
		...
	}

Of course, this assumes that atomic_t->counter underflows "correctly", just
like "unsigned int".

But again, do we really want this?

Oleg.