Re: 2.6.34-rc1: rcu lockdep bug?

["Américo Wang" <xiyou.wangcong@gmail.com>]

On Fri, Mar 12, 2010 at 02:03:19PM -0800, Paul E. McKenney wrote:
>On Fri, Mar 12, 2010 at 09:11:02PM +0800, Américo Wang wrote:
>> On Fri, Mar 12, 2010 at 7:11 PM, Eric Dumazet <eric.dumazet@gmail.com> wrote:
>> > Le vendredi 12 mars 2010 à 16:59 +0800, Américo Wang a écrit :
>> >> On Fri, Mar 12, 2010 at 4:07 PM, David Miller <davem@davemloft.net> wrote:
>> >> > From: Américo Wang <xiyou.wangcong@gmail.com>
>> >> > Date: Fri, 12 Mar 2010 15:56:03 +0800
>> >> >
>> >> >> Ok, after decoding the lockdep output, it looks like that
>> >> >> netif_receive_skb() should call rcu_read_lock_bh() instead of rcu_read_lock()?
>> >> >> But I don't know if all callers of netif_receive_skb() are in softirq context.
>> >> >
>> >> > Normally, netif_receive_skb() is invoked from softirq context.
>> >> >
>> >> > However, via netpoll it can be invoked essentially from any context.
>> >> >
>> >> > But, when this happens, the networking receive path makes amends such
>> >> > that this works fine.  That's what the netpoll_receive_skb() check in
>> >> > netif_receive_skb() is for.  That check makes it bail out early if the
>> >> > call to netif_receive_skb() is via a netpoll invocation.
>> >> >
>> >>
>> >> Oh, I see. This means we should call rcu_read_lock_bh() instead.
>> >> If Paul has no objections, I will send a patch for this.
>> >>
>> >
>> > Nope, its calling rcu_read_lock() from interrupt context and it should
>> > stay as is (we dont need to disable bh, this has a cpu cost)
>> >
>> 
>> Oh, but lockdep complains about rcu_read_lock(), it said
>> rcu_read_lock() can't be used in softirq context.
>> 
>> Am I missing something?
>
>Hmmm...  It is supposed to be OK to use rcu_read_lock() in pretty much
>any context, even NMI.  I will take a look.
>

Thanks! Please let me know if you have new progress.