On 12/13, Srivatsa S. Bhat wrote:

On 12/12/2012 11:32 PM, Oleg Nesterov wrote:

quoted

And _perhaps_ get_ can avoid it too?

I didn't really try to think, probably this is not right, but can't
something like this work?

	#define XXXX	(1 << 16)
	#define MASK	(XXXX -1)

	void get_online_cpus_atomic(void)
	{
		preempt_disable();

		// only for writer
		__this_cpu_add(reader_percpu_refcnt, XXXX);

		if (__this_cpu_read(reader_percpu_refcnt) & MASK) {
			__this_cpu_inc(reader_percpu_refcnt);
		} else {
			smp_wmb();
			if (writer_active()) {
				...
			}
		}

		__this_cpu_dec(reader_percpu_refcnt, XXXX);
	}

Sorry, may be I'm too blind to see, but I didn't understand the logic
of how the mask helps us avoid disabling interrupts..

Why do we need cli/sti at all? We should prevent the following race:

	- the writer already holds hotplug_rwlock, so get_ must not
	  succeed.

	- the new reader comes, it increments reader_percpu_refcnt,
	  but before it checks writer_active() ...

	- irq handler does get_online_cpus_atomic() and sees
	  reader_nested_percpu() == T, so it simply increments
	  reader_percpu_refcnt and succeeds.

OTOH, why do we need to increment reader_percpu_refcnt the counter
in advance? To ensure that either we see writer_active() or the
writer should see reader_percpu_refcnt != 0 (and that is why they
should write/read in reverse order).

The code above tries to avoid this race using the lower 16 bits
as a "nested-counter", and the upper bits to avoid the race with
the writer.

	// only for writer
	__this_cpu_add(reader_percpu_refcnt, XXXX);

If irq comes and does get_online_cpus_atomic(), it won't be confused
by __this_cpu_add(XXXX), it will check the lower bits and switch to
the "slow path".


But once again, so far I didn't really try to think. It is quite
possible I missed something.

Oleg.