Re: [PATCH] x86 MCE: shut up lockdep warning

[Shaohua Li <shaohua.li@intel.com>]

On Fri, May 08, 2009 at 05:37:57PM +0800, Andi Kleen wrote:
> >  - it works around a lockdep warning
> > 
> >  - you did not realize the real bug while the warning was plain
> 
> Well I'm not sure you understood it either :)  Actually I'm pretty
> sure you did not. It would have been good if you had awaited proper
> review comments before commiting your patch.
> 
> I don't think here's really a real bug because cpu hotunplug is single threaded
> anyways (cpu add remove lock) and we don't do multiple discoveries in parallel.
> 
> The reason the lock is there is only during bringup with multiple CPUs
> doing this in parallel on initial bootup. But we can't race against cpu 
> hotunplug there because there's not hotunplug before the system
> is up with all configured CPUs.
> 
> So I think any way of shutting up lockdep is fine here and your
> patch is overkill and disables interrupts unnecessarily. 
yes, this isn't a real bug, but just a false warning from lockdep. It's
my fault I didn't point this out first.

> >  - plus the patch introduces a fragile (because complex)
> >    work_on_cpu() call into the CPU hotplug path, which could have 
> >    caused followup regressions.
> 
> That's a reasonable point,  but doesn't seem strong enough to
> do full irq disabling. A better fix would be to find some other
> way to shut off this lockdep warning for this case. Is there
> such a way?
I tried to assign a lockdep class to the mce lock, which is the usual
way to avoid lockdep warning, but it appears not working here.

Thanks,
Shaohua