Date:	Tue, 18 May 2010 08:11:29 -0700
From:	"Paul E. McKenney" <paulmck@...ux.vnet.ibm.com>
To:	Mathieu Desnoyers <mathieu.desnoyers@...ymtl.ca>
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

On Tue, May 18, 2010 at 10:47:26AM -0400, Mathieu Desnoyers wrote:
> * Paul E. McKenney (paulmck@...ux.vnet.ibm.com) wrote:
> > On Mon, May 17, 2010 at 09:35:28PM -0400, Mathieu Desnoyers wrote:
> > > * 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:
> > 
> > [ . . . ]
> > 
> > > > > > 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. ;)
> > 
> > My first thought is that we have added quite a bit of RCU consistency
> > check code in the past few months, so we should see what bugs they find
> > and what bugs escape.  It is all too easy to create consistency check
> > code that is more trouble than it is worth.
> 
> Yes, although I expect that this new checking scheme will take some time
> to implement and mainline anyway (implementation effort which I might
> leave to someone else, as I have to focus on tracing at the moment).
> 
> > But in the meantime, let's see what would be required to check for
> > failures to insert grace-period delays:
> > 
> > o	There would need to be something like rcu_unreference(),
> > 	rcu_no_more_readers() or some such after the grace period.
> > 	The update side would then become something like the following:
> > 
> > 		oldp = rcu_dereference_protected(gp, &mylock);
> > 		rcu_assign_pointer(gp, newp);
> > 		synchronize_rcu();
> > 		rcu_no_more_readers(oldp);
> > 		kfree(oldp);
> 
> Replacing a kfree with a rcu_free(kfree, oldp) call that would include
> both could lessen the amount of typing:
> 
> #define rcu_free(freefct, ptr) \
> do { \
>   rcu_no_more_readers(ptr); \
>   freefct(ptr); \
> } while (0)

Or we could just rely on the existing debugobjects support that is
already in kfree().  ;-)

> > o	There would need to be something to check all of the pointers
> > 	traversed in the read-side critical sections:
> > 
> > 		rcu_read_lock();
> > 		...
> > 		p1 = rcu_dereference(gp1->field1);
> > 		...
> > 		p2 = rcu_dereference(gp2->field2);
> > 		...
> > 
> > 		rcu_validate(p1);
> > 		rcu_validate(p2);
> 
> Hrm, isn't the goal of this "rcu_validate(p1)" just to keep track of
> "p1" liveness ? Or do you plan to add a check there also ? I'm not sure
> I figure out what you are planning to validate here. I was thinking more
> in terms of
> 
>                 rcu_unreference(p1);
>                 rcu_unreference(p1);
> 
> that would be symmetric with the rcu_dereference.

My preference would be for people to just use the existing debugobjects
API, debug_check_no_obj_freed().  That is already in place, no need to
create RCU wrappers for it.

> > 		rcu_read_unlock();
> > 
> > One thing that bothers me about this is that we are forcing the developer
> > to do a lot of extra typing.  For example, rcu_no_more_readers() is in
> > a truth-and-beauty sense redundant with kfree() -- why type both?  The
> > same could be said about rcu_validate() and rcu_read_unlock(), but nested
> > RCU read-side critical sections make this difficult.
> 
> Ideally we'd like to add near-zero burden on developers, but I fear this
> cannot be done easily for read-side C.S.. As for write-side, we have to
> choose between tradeoff of genericity and less typing, e.g., between:
> 
>   rcu_free(kfree, ptr);
> and
>   rcu_kfree(ptr)
> 
> for the second, we would have to create a whole family of rcu_*free().
> 
> > 
> > Or am I misunderstanding what you are suggesting?
> 
> I'm only unsure about the "validate" part.

Again, we should just rely on the existing debugobjects function, letting
developers use it as they see fit.

							Thanx, Paul

> Thanks,
> 
> Mathieu
> 
> > 
> > 							Thanx, Paul
> 
> -- 
> 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