Message-ID: <20100518013528.GA21866@Krystal>
Date:	Mon, 17 May 2010 21:35:28 -0400
From:	Mathieu Desnoyers <mathieu.desnoyers@...ymtl.ca>
To:	"Paul E. McKenney" <paulmck@...ux.vnet.ibm.com>
Cc:	"Michael S. Tsirkin" <mst@...hat.com>,
	Peter Zijlstra <peterz@...radead.org>,
	linux-kernel@...r.kernel.org, mingo@...e.hu, laijs@...fujitsu.com,
	dipankar@...ibm.com, akpm@...ux-foundation.org,
	josh@...htriplett.org, dvhltc@...ibm.com, niv@...ibm.com,
	tglx@...utronix.de, rostedt@...dmis.org, Valdis.Kletnieks@...edu,
	dhowells@...hat.com, eric.dumazet@...il.com,
	Arnd Bergmann <arnd@...ay.de.ibm.com>,
	Arnd Bergmann <arnd@...db.de>
Subject: Re: [PATCH RFC tip/core/rcu 23/23] vhost: add __rcu annotations

* Paul E. McKenney (paulmck@...ux.vnet.ibm.com) wrote:
> On Mon, May 17, 2010 at 07:40:25PM -0400, Mathieu Desnoyers wrote:
> > * Paul E. McKenney (paulmck@...ux.vnet.ibm.com) wrote:
> > > On Mon, May 17, 2010 at 06:00:25PM -0400, Mathieu Desnoyers wrote:
> > > > * Paul E. McKenney (paulmck@...ux.vnet.ibm.com) wrote:
> > > > > On Mon, May 17, 2010 at 11:33:49PM +0300, Michael S. Tsirkin wrote:
> > > > > > On Thu, May 13, 2010 at 08:23:40AM -0700, Paul E. McKenney wrote:
> > > > > > > On Thu, May 13, 2010 at 03:07:23PM +0200, Peter Zijlstra wrote:
> > > > > > > > On Wed, 2010-05-12 at 16:00 -0700, Paul E. McKenney wrote:
> > > > > > > > > Any thoughts?  One approach would be to create a separate lockdep class
> > > > > > > > > for vhost workqueue state, similar to the approach used in instrument
> > > > > > > > > rcu_read_lock() and friends. 
> > > > > > > > 
> > > > > > > > workqueue_struct::lockdep_map, its held while executing worklets.
> > > > > > > > 
> > > > > > > > lock_is_held(&vhost_workqueue->lockdep_map), should do as you want.
> > > > > > > 
> > > > > > > Thank you, Peter!!!
> > > > > > > 
> > > > > > > 							Thanx, Paul
> > > > > > 
> > > > > > vhost in fact does flush_work rather than
> > > > > > flush_workqueue, so while for now everything runs
> > > > > > from vhost_workqueue in theory nothing would break
> > > > > > if we use some other workqueue or even a combination
> > > > > > thereof.
> > > > > > 
> > > > > > I guess when/if this happens, we could start by converting
> > > > > > to _raw and then devise a solution.
> > > > > 
> > > > > If there are a small finite number of work queues involved, we can
> > > > > easily do something like:
> > > > > 
> > > > > 	#ifdef CONFIG_PROVE_RCU
> > > > > 	int in_vhost_workqueue(void)
> > > > > 	{
> > > > > 		return in_workqueue_context(vhost_workqueue) ||
> > > > > 		       in_workqueue_context(vhost_other_workqueue) ||
> > > > > 		       in_workqueue_context(yet_another_vhost_workqueue);
> > > > > 	}
> > > > > 	#endif /* #ifdef CONFIG_PROVE_RCU */
> > > > > 
> > > > > Seem reasonable?
> > > > > 
> > > > > > By the way what would be really nice is if we had a way
> > > > > > to trap when rcu protected pointer is freed without a flush
> > > > > > while some reader is running. Current annotation does not
> > > > > > allow this, does it?
> > > > > 
> > > > > Right now, it does not, but I wonder if something like Thomas's and
> > > > > Mathieu's debugobjects work could be brought to bear on this problem?
> > > > > This would need to be implemented in vhost, as synchronize_rcu() has
> > > > > no way to know what memory it is flushing, nor does flush_work().
> > > > 
> > > > We can think of my recent debugobjects addition as a small state machine
> > > > that is described by the code that owns the objects. At each state
> > > > transition, the code passes the expected state as well as the next
> > > > state.
> > > > 
> > > > The current implementation can only keep track of a single "state" per
> > > > object at once. This should be extended to be able to count the number
> > > > RCU read side C.S. in flight that are accessing to an object.
> > > 
> > > Not a problem, as vhost doesn't use call_rcu().  So there won't be a
> > > conflict between different debugobjects views of the same memory.
> > 
> > Not quite sure I follow you here.
> 
> vhost uses only synchronize_rcu() and flush_work().  The existing
> debugobjects tagging would therefore be unaware of the actual object,
> instead tagging the rcu_head that synchronize_rcu() allocated on
> the stack, and being out of the picture completely in the case of
> flush_work().
> 
> Either way, RCU is completely unaware of exactly which structure is
> being pushed through a grace period, so RCU's debugobjects tagging cannot
> possibly conflict with any tagging that vhost does.

I see.

> 
> > > > We could use a hook in rcu_dereference (which knows about the object)
> > > > and a hook in rcu_read_unlock (which determines the end of valid object
> > > > use).
> > > > 
> > > > We should hook into rcu_assign_pointer() to detect RCU structure
> > > > privatization. It should put these objects in a "privatized" hash table.
> > > > 
> > > > We should also hook into synchronize_rcu/sched() to remove the
> > > > privatized structures from the privatized hash.
> > > > 
> > > > A hook in "kfree" (maybe a new rcu_free(void (fctptr*)(void *)) wrapper ?)
> > > > would call a debugobject hook that would lookup the "privatized" hash.
> > > > If it contains the object to free, we check if there are RCU read-side
> > > > C.S. in flight using this object at the same time, and show an error if
> > > > both are true.
> > > 
> > > I believe that we can't bury this into the RCU primitives, because
> > > rcu_read_unlock() doesn't know what objects were referenced in the
> > > RCU read-side critical section.
> > 
> > Well, if we can find a way to match a sequence of rcu_dereference
> > performed from a thread with the following rcu_read_unlock(), then we
> > might have the information we need. But we would have to somehow tie the
> > debugobject context to the thread context. That sounds too complex for
> > what we are trying to achieve here.
> 
> Indeed!  Especially given the fact that RCU read-side critical sections
> can be nested.  Which rcu_dereference() calls go with which RCU read-side
> critical section?

Good point.

> 
> > > But perhaps we should be simply treating this as a use-after-free
> > > problem, so that RCU is not directly involved. Isn't that the standard
> > > use of debugobjects anyway?
> > 
> > OK so we could tie "rcu_dereference" do debugobjects, and free would be
> > a standard free. Yes, I think it could be done. It looks a bit like the
> > memory allocation debugging code. If we know that a certain
> > rcu_dereference always access dynamically allocated memory, we could
> > probably add some checks there based on the memory allocator debug
> > objects.
> 
> We probably need vhost to add code at the end of the relevant RCU
> read-side critical section checking that the pointers returned by
> any rcu_dereference() calls still point to valid memory.  Don't get
> me wrong, your approach could find bugs in which someone forgot to
> remove the RCU-protected structure from a public list, but it could
> not detect failure to wait a grace period between the time of removal
> and the time of freeing.

Good point too. So something like a new rcu_unreference() (or feel free
to find any better name) ;) that would be compiled out normally, but
would call into debugobjects might do the trick. We would have to add
these annotations to match every rcu_dereference() though, might means a
lot of new lines of code. On the plus side, that looks like a good audit
of RCU read-side use. ;)

Thanks,

Mathieu

> 
> 							Thanx, Paul
> 
> > Thanks,
> > 
> > Mathieu
> > 
> > 
> > > 
> > > 							Thanx, Paul
> > > 
> > > > Thoughts ?
> > > > 
> > > > Mathieu
> > > > 
> > > > 
> > > > > 
> > > > > 						Thanx, Paul
> > > > 
> > > > -- 
> > > > Mathieu Desnoyers
> > > > Operating System Efficiency R&D Consultant
> > > > EfficiOS Inc.
> > > > http://www.efficios.com
> > 
> > -- 
> > Mathieu Desnoyers
> > Operating System Efficiency R&D Consultant
> > EfficiOS Inc.
> > http://www.efficios.com

-- 
Mathieu Desnoyers
Operating System Efficiency R&D Consultant
EfficiOS Inc.
http://www.efficios.com
--
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