Date:	Wed, 27 May 2015 22:41:37 -0700
From:	"Nicholas A. Bellinger" <nab@...ux-iscsi.org>
To:	paulmck@...ux.vnet.ibm.com
Cc:	Bart Van Assche <bart.vanassche@...disk.com>,
	"Nicholas A. Bellinger" <nab@...erainc.com>,
	target-devel <target-devel@...r.kernel.org>,
	linux-scsi <linux-scsi@...r.kernel.org>,
	linux-kernel <linux-kernel@...r.kernel.org>,
	Christoph Hellwig <hch@....de>, Hannes Reinecke <hare@...e.de>,
	Sagi Grimberg <sagig@...lanox.com>
Subject: Re: [PATCH-v2 0/4] target: Eliminate se_port +
 t10_alua_tg_pt_gp_member

On Wed, 2015-05-27 at 13:36 -0700, Paul E. McKenney wrote:
> On Tue, May 26, 2015 at 10:13:02PM -0700, Nicholas A. Bellinger wrote:
> > On Tue, 2015-05-26 at 14:44 +0200, Bart Van Assche wrote:
> > > On 05/26/15 08:57, Nicholas A. Bellinger wrote:
> > > >    - Add various rcu_dereference and lockless_dereference RCU notation
> > > 
> > > Hello Nic,
> > > 
> > > Feedback from an RCU expert (which I'm not) would be appreciated here. 
> > > But my understanding is that lockless_dereference(p) should be used for 
> > > a pointer p that has *not* been annotated as an RCU pointer. I think in 
> > > the for-next branch of the target repository that this macro is used to 
> > > access RCU-annotated pointers. Is that why sparse complains about how 
> > > lockless_dereference() is used in the target tree ?
> > > 
> > 
> > Was curious about this myself..  Thanks for raising the question!
> > 
> > The intention of lockless_dereference() in both this and preceding
> > series is for __rcu protected pointers that are accessed outside of
> > rcu_read_lock() protection, and who's lifetime is controlled by a:
> > 
> >   - struct kref
> >   - struct percpu_ref
> >   - struct config_group symlink
> >   - RCU updater path with some manner of mutex or spinlock held
> > 
> > This is supposed to be following Paul's comment in rcupdate.h:
> > 
> >  * Similar to rcu_dereference(), but for situations where the pointed-to
> >  * object's lifetime is managed by something other than RCU.  That
> >  * "something other" might be reference counting or simple immortality.
> > 
> > Paul, would you be to kind to clarify the intention for us..?
> 
> The lockless_dereference() primitive is to be used for pointers that
> are -not- marked with __rcu.  In fact, the sparse tool should yell
> at you if you use lockless_dereference() on an __rcu-marked pointer.

Yep, definitely wrong usage of lockless_dereference on my part.

Thanks for the clarification.

> You could use smp_store_release() to update the pointer when inserting
> new data.  If you are using one of the lists, then the _rcu variant of the
> list-insert macro should be used (list_add_rcu()), because that is needed
> to make sure that the reader sees a properly initialized new element.
> 
> If you have a pointer that is sometimes protected by RCU and other times
> protected by something else, you still use one of the rcu_dereference()
> macros to access it.  For example, if a given RCU-protected pointer is
> protected either by RCU or by some lock, you might write common code
> that is called from either context as follows:
> 
> 	p = rcu_dereference_check(pointer, lockdep_is_held(&some_lock));
> 
> Does that help, or am I missing your point?
> 

This makes more sense now.

Ok, so for an updater path where a __rcu protected pointer is being
dereferenced with a lock held synchronizing modification of an
hlist_head or hlist_node, the rcu_dereference_check() usage is clear to
me..

What I'm still unclear about is other three cases above, where a __rcu
protected pointer is dereferenced outside of the updater path, but it's
release is protected by some external means; kref, percpu_ref, or a
configfs parent/child config_group reference.

For example, say a __rcu protected pointer is dereferenced under
rcu_read_lock().  The data structure itself contains a percpu_ref that
is incremented under rcu_read_lock(), and also contains a rcu_head.
rcu_read_unlock() happens immediately after the percpu_ref has been
incremented.

If this structure is then attempted to be released from an updater path,
it first blocks on a completion waiting for the percpu_ref count to
return to zero, before invoking the final kfree_rcu().

So the question I'm getting at is, what's the correct notation for
dereferencing a __rcu pointer outside of rcu_read_lock(), who's data
structure is protected by some manner of reference counting obtained
under rcu_read_lock(), that prevents kfree_rcu() from happening until
the reference count is dropped..?

--nab

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

Powered by blists - more mailing lists