Re: [PATCH] slub: fix slab_pad_check()

["Paul E. McKenney" <paulmck@linux.vnet.ibm.com>]

On Thu, Sep 03, 2009 at 05:44:54PM -0500, Christoph Lameter wrote:
> On Thu, 3 Sep 2009, Eric Dumazet wrote:
> 
> > It seems very smart, but needs review of all callers to make sure no slabs
> > are waiting for final freeing in call_rcu queue on some cpu.
> 
> Yes. Again this is the first time we encounter a situation where a
> DESTROY_BY_RCU slab has to be destroyed. So the review is quite short.
> 
> > I suspect most of them will then have to use rcu_barrier() before calling
> > kmem_cache_destroy(), so why not factorizing code in one place ?
> 
> There are different forms of RCU which require different forms of
> barriers. Its best to leave that up to the user. Again the user must make
> sure that no objects are in use before a slab is destroyed. For
> SLAB_DESTROY_BY_RCU this means that there are no potential outstanding
> reads on the structure. You may need an rcu_barrier() to accomplish that.
> 
> Slight variations in the use of RCU could require different method. Better
> reduce the entanglement of slabs to RCU to a mininum possible.

If it were the user of the slab who was invoking some variant of
call_rcu(), then I would agree with you.

However, call_rcu() is instead being invoked by the slab itself in the
case of SLAB_DESTROY_BY_RCU, so that there is no variation in usage.
Requiring that the user call rcu_barrier() is asking for subtle bugs.
Therefore, the best approach is to have kmem_cache_destroy() handle
the RCU cleanup, given that this cleanup is for actions taken by
kmem_cache_free(), not by the user.

							Thanx, Paul