Re: [PATCH/RFC] remove irqs_disabled warning from local_bh_enable

[Peter Zijlstra <peterz@infradead.org>]

On Tue, 2008-06-17 at 16:55 -0700, Linus Torvalds wrote:
> 
> On Tue, 17 Jun 2008, Johannes Berg wrote:
> >
> > This warning has started to trigger with mac80211 because it can, under
> > some circumstances, use spin_lock_bh() protected sections within
> > irq-disabled sections. Is that a bug?
> 
> Yes, it's a bug.
> 
> Why? Not because of the "spin_lock_bh()" itself, but because of the 
> _unlock_, which does a "local_bh_enable_ip()", which in turn will check 
> the whole "do_softirq()" if it was the last softirq_count.
> 
> And you must not do softirq's when hard-irq's were disabled!
> 
> So it should in theory be ok (but perhaps a bit odd) to do something like
> 
> 	spin_lock_irq(&irq_lock);
> 	..do something..
> 	spin_lock_bh(&bh_lock);
> 	spin_unlock_irq(&irq_lock);
> 	.. do something else ..
> 	spin_unlock_bh(&bh_lock);
> 
> where the "spin_lock_bh()" itself is in an irq-locked context - as long as 
> the "spin_unlock_bh()" is *not*.

I would suggest discouraging such madne^Wcreativity, its gains are
dubious at best and it doesn't make the locking any more obvious and
could be an indication of messy locking to begin with.

So I would like to see Johannes' other patch that allows all of us to
enjoy the warning he ran into ;-)

Peter