Re: [PATCH 2/3] rcu: Create rcu_sync infrastructure

[Oleg Nesterov <oleg@redhat.com>]

On 10/04, Peter Zijlstra wrote:
>
> On Fri, Oct 04, 2013 at 01:15:13PM +0200, Oleg Nesterov wrote:
> > Of course we can fix this, but wait_for_completion/complete already
> > does the necessary work: x->done acts as a resource counter which is
> > always checked/incremented/decremented under the same lock.
>
> Urgh, so now you're not using a semaphore as completion but using a
> completion as semaphore.

Yes. And we can use it as semaphore only because we already have gp_count.

Of course, we could add ->gp_mutex and change enter/exit to lock/unlock
it if exclusive. But I do not really like it even if this can makes the
code a bit (only a little bit) simpler. This will increase sizeof() and
for no reason, I think, because to some degree this will also duplicate
the synchronization logic. And while this doesn't really matter, this
will penalize the contending writers a bit: it is not easy (but of
course possible) to take this mutex after rss->gp_count++, and we do
want to increment this counter "asap".

Oleg.