Re: [Bugme-new] [Bug 10326] New: inconsistent lock state in net_rx_action

[Peter Zijlstra <a.p.zijlstra@chello.nl>]

On Thu, 2008-03-27 at 02:18 -0700, Andrew Morton wrote:
> On Thu, 27 Mar 2008 09:55:42 +0100 Jarek Poplawski <jarkao2@gmail.com> wrote:
> 
> > On Wed, Mar 26, 2008 at 05:14:03PM -0700, Andrew Morton wrote:
> > ...
> > > > >> http://bugzilla.kernel.org/show_bug.cgi?id=10326
> > ...
> > > No, it's not an irq_disable() thing, directly.
> > > 
> > > What lockdep is saying is that sky2_poll() is taking napi->poll_lock for
> > > writing with softirqs enabled, but net_rx_action() takes the same lock from
> > > within softirq context.
> > > 
> > > If sky2_poll() always takes napi->poll_lock under local_irq_disable() then
> > > that would be a lockdep bug.
> > 
> > sky2_poll() doesn't take napi->poll_lock; this lock is taken by
> > netpoll_poll() before calling sky2_poll(). And before this hardirqs
> > are disabled in write_msg(). So, theoretically lockdep could be right
> > if sky2_poll() would enable irqs after this. (If it were done in
> > netpoll - lockdep should warn before or after sky2_poll() call.)
> > But I really can't see any such possibility in sky2_poll().
> 
> I can't spot it from a five-minute read either.  gcc's autoinlining really
> makes this sort of thing much harder than it used to be :(
> 
> Anyway, the accusation is that lockdep is busted, in that it doesn't realise that
> local_irq_disable() blocks softirqs.
> 
> I bet the net code is wrong and we missed it ;)

How about this:

<irqs disabled>

  netpoll_poll()
    poll_napi()
      spin_trylock(&napi->poll_lock)
      poll_one_napi()
        napi->poll() := sky2_poll()
          napi_complete()
            local_irq_disable()
            local_irq_enable() <--- *BUG*

  <irq>
    irq_exit()
      do_softirq()
        net_rx_action()
          spin_lock(&napi->poll_lock) <--- Deadlock!

Because we still hold the lock....