On Sat, 2006-07-29 at 03:40 +0200, Patrick McHardy wrote:

Martin Josefsson wrote:

quoted

As a somewhat related note, I've just digged a bit through RCU land,
talked to dipankar and mckenney, and discovered that rcu_read_lock() /
rcu_read_unlock() aren't strictly needed in softirqs since preempt is
already disabled in softirqs. This means that you can use the result of
the rcu read-side critical outside of the rcu_read_lock() /

Thats true, but in this case the code is executed both in softirq-
and user-context. Using rcu_read_lock and still relying on softirq
properties outside the locked section is also very confusing in my
opinion.

Yes it is very fishy relying on the softirq properties, especially since
they don't apply to -rt kernels and that there might be other changes in
this area in the future. It's not recommended.

-- 
/Martin