Re: [PATCH RFC] perf: fix find_swevent_head() RCU lockdep splat

From: Paul E. McKenney
Date: Thu May 13 2010 - 15:46:15 EST


On Thu, May 13, 2010 at 09:03:28PM +0200, Frederic Weisbecker wrote:
> On Thu, May 13, 2010 at 11:25:56AM -0700, Paul E. McKenney wrote:
> > This commit guesses at the perf_cpu_context locking design and deploys
> > an rcu_dereference_check() accordingly. The design appears to require
> > that a given CPU be accessing its own per_cpu_context or that it be
> > traversing under RCU protection.
> >
> > Signed-off-by: Paul E. McKenney <paulmck@xxxxxxxxxxxxxxxxxx>
> > Cc: Frederic Weisbecker <fweisbec@xxxxxxxxx>
> >
> > perf_event.c | 4 +++-
> > 1 file changed, 3 insertions(+), 1 deletion(-)
> >
> > diff --git a/kernel/perf_event.c b/kernel/perf_event.c
> > index a4fa381..002791c 100644
> > --- a/kernel/perf_event.c
> > +++ b/kernel/perf_event.c
> > @@ -4074,7 +4074,9 @@ find_swevent_head(struct perf_cpu_context *ctx, u64 type, u32 event_id)
> >
> > hash = swevent_hash(type, event_id);
> >
> > - hlist = rcu_dereference(ctx->swevent_hlist);
> > + hlist = rcu_dereference_check(ctx->swevent_hlist,
> > + rcu_read_lock_held() ||
> > + ctx == &__get_cpu_var(perf_cpu_context));
> > if (!hlist)
> > return NULL;
> >
>
>
> Hmm, that's not exactly that. It will always be the ctx of this cpu
> but not always under rcu read lock. I mean touching the current cpu
> ctx is not inherently safe.
>
> In fact we have two paths:
>
> perf_swevent_enable() gets the hlist and if it is called it means
> that this hlist is not supposed to be NULL. If it is, it's a bug.
>
> If we have created a software event, the hlist has been allocated
> and perf_swevent_enable() is called later to activate this event.
> May be I shouldn't use rcu_dereference() here but a simple dereference.
> And the hlist can't be freed under us at this time so we don't need
> rcu_read_lock().
>
> OTOH, do_perf_sw_event() can be called anytime so it need this
> rcu_read_lock().
>
>
> On the perf_swevent_enable() path, what prevents the hlist to be
> freed under us is the ctx->lock. Because we won't ever remove
> an event from its context list outside this lock, and we might only
> release the hlist after a software event gets removed from its
> context list.
>
> So either we do this:
>
> hlist = rcu_dereference_check(ctx->swevent_hlist,
> rcu_read_lock_held() ||
> raw_spin_lock_is_held(&ctx->lock));

Something very similar to the above was in fact my first guess, but as
Ingo can attest, this results in build errors. The problem is that
perf_cpu_context does not have a ->lock field. Hmmm... But it does
contain a struct perf_event_context (ctx) and a pointer to another
(task_ctx). So, would one of the following work?

list = rcu_dereference_check(ctx->swevent_hlist,
rcu_read_lock_held() ||
lockdep_is_held(&ctx->ctx.lock));

list = rcu_dereference_check(ctx->swevent_hlist,
rcu_read_lock_held() ||
lockdep_is_held(&ctx->task_ctx->lock));

If the latter, can ->task_ctx ever be NULL, and if so, what should
I do then?

> or:
>
>
> hlist = ctx->swevent_hlist;

This will be flagged as an error by Arnd's sparse-based checking. :-(

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