Re: Semantics of smp_mb() [was : Re: [PATCH] Fix RCU race in access of nohz_cpu_mask ]

[Oleg Nesterov <oleg@tv-sign.ru>]

Srivatsa Vaddagiri wrote:
> 
> Oleg, with all these inputs, I consider the patch I had sent to be correct.

Yes, I think so. My suspects were due to my misunderstanding.
I was wrong when I said that we can ignore rmb() part of mb()
in case when start_hz_timer() runs after rcu_start_batch().

	rcp->cur++;
	// ->cur will be visible to other cpus
	// _before_ we will *READ* nohz_cpu_mask.
	// we don't have any 'timing' problems.
	// In other words: if another cpu does not
	// see the new value - we did not read this
	// mask yet.
	smp_mb();

Thanks for your patience.

> P.S :- Thanks to everybody who reponded clarifying this subject.

Yes! it was really helpful, thanks to all. I think it would be great
to have Paul's very clear (and short!) explanation somewhere in
Documentation/.

Oleg.