Re: [PATCH v2 10/10] nf_conntrack: Use rcu_barrier() and fix kmem_cache_create flags

[Patrick McHardy <kaber@trash.net>]

Jesper Dangaard Brouer wrote:
> On Wed, 2009-06-24 at 15:58 +0200, Patrick McHardy wrote:
>> Jesper Dangaard Brouer wrote:
>>> Adjusting SLAB_DESTROY_BY_RCU flags.
>>>
>>>  kmem_cache_create("nf_conntrack", ...) does not need the
>>>  SLAB_DESTROY_BY_RCU flag.
>> It does need it. We're using it instead of call_rcu() for conntracks.
>>
>>>  But the
>>>  kmem_cache_create("nf_conntrack_expect", ...) should use the
>>>  SLAB_DESTROY_BY_RCU flag, because it uses a call_rcu() callback to
>>>  invoke kmem_cache_free().
>> No, using call_rcu() means we don't need SLAB_DESTROY_BY_RCU.
>> Please see the note in include/linux/slab.h.
> 
> Oh, I see.  The description is some what cryptic, but I think I got it,
> after reading through the code.
> 
> BUT this still means that we need to do rcu_barrier() if the
> SLAB_DESTROY_BY_RCU is NOT set and we do call_rcu() our self.

Correct, in that case its necessary.

> My understanding for the code is (please feel free to correct me): that
> if SLAB_DESTROY_BY_RCU _is_ set, then the __cache_shrink() call will
> call drain_freelist(), which calls slab_destroy().
> 
> If SLAB_DESTROY_BY_RCU _is_ set, then slab_destroy() will then start a
> call_rcu() callback to kmem_rcu_free() which calls kmem_cache_free().
> Given that the callback code kmem_rcu_free() is not removed, we are not
> worried about unloading the module at this point.

Yep, thats my understanding as well.

> I'm a bit worried about what happens if __kmem_cache_destroy() is
> invoked and there is still callbacks for kmem_rcu_free() in flight?
> The synchronize_rcu() between __cache_shrink() and
> __kmem_cache_destroy() should perhaps be changed to rcu_barrier()?
> 
> But I'm sure that the SLAB/MM guys will tell me that this case is
> handled (and something about its unlinked from the appropiate
> lists)??? ;-)

I'll leave that question to the MM guys :)

>>> RCU barriers, rcu_barrier(), is inserted two places.
>>>
>>>  In nf_conntrack_expect.c nf_conntrack_expect_fini() before the
>>>  kmem_cache_destroy(), even though the use of the SLAB_DESTROY_BY_RCU
>>>  flag, because slub does not (currently) handle rcu sync correctly.
>> I think that should be fixed in slub then.
> 
> I don't think so, we/I'm are talking about "nf_conntrack_expect" and not
> "nf_conntrack" slab.  Clearly the slab "nf_conntrack" is handled
> correcly (according to description above). 
> 
> We still need to make sure the callbacks for "nf_conntrack_expect", are
> done before unloading/removing the code they are about to call.

Yes, my response was referring to potential sl*b bugs, but
you're correct, we do need rcu_barrier() for expectations.

>>>  And in nf_conntrack_extend.c nf_ct_extend_unregister(), inorder to
>>>  wait for completion of callbacks to __nf_ct_ext_free_rcu(), which is
>>>  invoked by __nf_ct_ext_add().  It might be more efficient to call
>>>  rcu_barrier() in nf_conntrack_core.c nf_conntrack_cleanup_net(), but
>>>  thats make it more difficult to read the code (as the callback code
>>>  in located in nf_conntrack_extend.c).
>> This one looks fine.
> 
> Should I make two different patchs?

Either way is fine.