Re: lockdep: possible reason: unannotated irqs-off. (was: Re: Linux 2.6.38-rc4)

[Peter Zijlstra <peterz@infradead.org>]

On Tue, 2011-02-08 at 11:17 +0100, Borislav Petkov wrote:
> 
> So, I'm getting the warning below early in the boot process. And yes, I
> didn't have it on -rc3. 2.6.38-rc4-00001-g1e554e3-dirty means a debug
> diff ontop of -rc4 which shouldn't have anything to do with this splat
> since all it does is a couple of printk's due to -rc3 not suspending to
> disk properly in some cases.
> 
> Now, I'm not going to even pretend to understand the code but here's
> what I can read out, you tell me whether it makes sense.
> 
> spawn_ksoftirqd() is one of the early initcalls that gets called and
> it's notifier callback does kthread_bind() and you can follow in the
> backtrace below that this thing comes down to del_timer_sync() which
> does the lockdep annotation. Now, problem as I see it, is that hardirqs
> were disabled when we were called although it doesn't say so in the
> irqtrace events dump after the calltrace: "hardirqs last enabled at
> (999)" and our irq event stamp is 1000.
> 
> There are actually at least three del_timer_sync()'s inflight so the
> problem could be there somewhere, I dunno.
> 
> So, is it a wrong lockdep annotation or is it a real problem? I've
> attached dmesg and config.

Argh! Its an annotation nightmare that.. it didn't trigger for me when
running that because I didn't have DEBUG_LOCKDEP=y.

OK, let me try and come up with another way to annotate this
del_timer_sync() muck, the trouble is we want that lock to be called
with BH disabled, but simply doing local_bh_disable()/local_bh_enable()
has the nasty side effect of calling __do_softirq(). Faking IRQ state
will trip this check_flags() debug muck..

We used to have local_irq_disable()/local_irq_enable() around it, but
then people wanted to use del_timer_sync() from softirq context..