Messages in this thread | | | Date | Fri, 4 Oct 2013 21:06:53 +0200 | From | Oleg Nesterov <> | Subject | Re: [PATCH 2/3] rcu: Create rcu_sync infrastructure |
| |
On 10/04, Peter Zijlstra wrote: > > On Fri, Oct 04, 2013 at 05:13:23PM +0200, Oleg Nesterov wrote: > > > I am not trying saying this feature is "must have", of course it > > is not. The only problem, I am a bit puzzled why you dislike it > > that much. > > The reason I dislike it is because I feel we're now mixing two objects > into one; one object doing mutual exclusion and one object being > terribly smart with sync_rcu.
OK, I see your point.
But rcu_sync_struct has to serialize the writers anyway. The only question is how many other writers the thread doing ->sync() should wakeup and when.
And otoh. Currently nobody needs the non-exclusive mode (cpu-hotplug doesn't care because it is always exclusive itself). And in fact you initially argued with wake_up_all ;) "exclusive" is more natural, it is like rw_semaphore.
However, yes-yes-yes, I do think that we need the non-exclusive mode too, at least for percpu_down_write_nonexclusive() which I think we need as well.
Oleg.
| |