Re: [PATCH 1/2] trace_events_filter: use rcu_assign_pointer() when setting ftrace_event_call->filter

[Tejun Heo <tj@kernel.org>]

On Wed, Nov 23, 2011 at 10:16:47AM -0500, Steven Rostedt wrote:
> [ Added Paul to Cc ]
> 
> On Tue, 2011-11-22 at 17:46 -0800, Tejun Heo wrote:
> > ftrace_event_call->filter is sched RCU protected but didn't use
> > rcu_assign_pointer().  Fix it.
> 
> Is it really needed? Maybe just for documentation but I'm not sure this
> use is required because all use cases have synchronize_sched() used,
> which is a big hammer compared to the rcu_assign_pointer().

Oh yeah, I think we do.  synchronize_sched() is to drain users of the
old pointer.  Whether synchronize_sched() or call_rcu() is used is
irrelevant to the synchronization of new pointer.

rcu_assign_pointer() is to synchronize against rcu_dereference()
dereferencing the new one.  More specifically, we need it for the
writer memory barrier, so that it matches the data dependency read
barrier in rcu_dereference() when it access the new pointer;
otherwise, it may fetch the old values from before the new area is
initialized on archs where data dependency barrier isn't noop.

> We update filter here, and then call synchronize_sched() before we free
> the filter_item->filter.
...
> Again you can see that synchronize_sched() is called here.

So, synchronized_sched() being called after isn't relevant.  We want
smp_wmb() between data structure initialization and assignment of the
new pointer.

Thank you.

-- 
tejun