Re: [PATCH 2/9] slab: remove synchronous rcu_barrier() call in memcg cache release path

[Joonsoo Kim <iamjoonsoo.kim-Hm3cg6mZ9cc@public.gmane.org>]

On Sat, Jan 14, 2017 at 10:19:21AM -0500, Tejun Heo wrote:
> Hello, Vladimir.
> 
> On Sat, Jan 14, 2017 at 04:19:39PM +0300, Vladimir Davydov wrote:
> > On Sat, Jan 14, 2017 at 12:54:42AM -0500, Tejun Heo wrote:
> > > This patch updates the cache release path so that it simply uses
> > > call_rcu() instead of the synchronous rcu_barrier() + custom batching.
> > > This doesn't cost more while being logically simpler and way more
> > > scalable.
> > 
> > The point of rcu_barrier() is to wait until all rcu calls freeing slabs
> > from the cache being destroyed are over (rcu_free_slab, kmem_rcu_free).
> > I'm not sure if call_rcu() guarantees that for all rcu implementations
> > too. If it did, why would we need rcu_barrier() at all?
> 
> Yeah, I had a similar question and scanned its users briefly.  Looks
> like it's used in combination with ctors so that its users can
> opportunistically dereference objects and e.g. check ids / state /
> whatever without worrying about the objects' lifetimes.

Hello, Tejun.

Long time no see! :)

IIUC, rcu_barrier() here prevents to destruct the kmem_cache until all
slab pages in it are freed. These slab pages are freed through call_rcu().

Your patch changes it to another call_rcu() and, I think, if sequence of
executing rcu callbacks is the same with sequence of adding rcu
callbacks, it would work. However, I'm not sure that it is
guaranteed by RCU API. Am I missing something?

Thanks.