Re: Regression - locking (all from 2.6.28)

[Peter Zijlstra]

On Mon, 2009-03-02 at 12:11 -0800, Andrew Morton wrote:

> > Mar  1 00:07:03 localhost kernel: [   86.440261] =========================================================
> > Mar  1 00:07:03 localhost kernel: [   86.440266] [ INFO: possible irq lock inversion dependency detected ]
> > Mar  1 00:07:03 localhost kernel: [   86.440271] 2.6.29-rc6-mm1-hanny #17
> > Mar  1 00:07:03 localhost kernel: [   86.440273] ---------------------------------------------------------
> 
> I stared at this for a while, but my brain broke trying to work out
> what lockdep is trying to tell us.
> 
> > Mar  1 00:07:03 localhost kernel: [   86.440277] Xorg/2733 just changed the state of lock:
> > Mar  1 00:07:03 localhost kernel: [   86.440280]  (fasync_lock){.-....}, at: [<c01952bb>] kill_fasync+0x20/0x3a
> > Mar  1 00:07:03 localhost kernel: [   86.440292] but this lock took another, HARDIRQ-READ-irq-unsafe lock in the past:
> > Mar  1 00:07:03 localhost kernel: [   86.440296]  (&f->f_lock){+.+...}
> 
> This message needs help.  A lock cannot "take" another lock. 

It seemed a simple enough way to tell that the latter lock nests inside
the former lock.

So what its saying is that we have:

  fasync_lock
    f->f_lock

nesting, and fasync_lock got used in hardirq context, but the lock that
was previously found to nest inside, was an IRQ-unsafe lock.

So $CODE code take f->f_lock, then IRQ could happen and fasync_lock,
f->f_lock could happen and we'd be stuck.

Would something like:

"but this lock had a %s-irq-unsafe nestee in the past:" read better?

> And why
> is f_lock described as "HARDIRQ-READ-irq-unsafe"?  It's a spinlock and
> the "READ" part is not relevant.

I think that's a bug due to the recent irq state tracking generalization
patches, will hunt.

> > Mar  1 00:07:03 localhost kernel: [   86.440299] 
> > Mar  1 00:07:03 localhost kernel: [   86.440300] and interrupts could create inverse lock ordering between them.
> > Mar  1 00:07:03 localhost kernel: [   86.440302] 



--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/