Re: processes hung after sys_renameat, and 'missing' processes

[Peter Zijlstra <peterz@infradead.org>]

On Sun, 2012-06-03 at 17:16 -0700, Linus Torvalds wrote:
> On Sun, Jun 3, 2012 at 5:00 PM, Dave Jones <davej@redhat.com> wrote:
> >
> > sysrq-d: http://fpaste.org/ow9O/
> 
> Ugh. I'm adding PeterZ to the cc, just to see if he can make more sense of it.
> 
> Peter, is there no way to make the lock thing print not just the lock
> class name, but also the pointer to the actual *instance* of the lock
> held?

Sorta, we have a pointer to the struct lockdep_map inside whatever lock
type. But we don't have the lock type so we cannot actually provide the
pointer to the spinlock_t struct mutex etc..

> Also, it's a bit unclear to me, but I *think* that most of those users
> don't actually "hold" the lock - they are waiting for it. Yes/no?

Yes, this is waiting to acquire, lockdep started out with 2 hooks, one
before the actual acquire and one on release. Its done before the actual
acquire so we can warn before we lock up in case of an actual deadlock.

>  Does
> the lockdep information have the capability to distinguish between
> "waiting for" vs "actually successfully owns the lock"?

Sometimes.. when build with lockstat we have enough hooks to do this. I
suppose I could make all those hooks available for all of lockdep and
track the per lock state more accurate in order to improve this
printout.