Re: lockdep whine in 2.6.26-rc2-mm1

[Matthew Wilcox <matthew@wil.cx>]

On Wed, May 14, 2008 at 08:56:39AM -0700, Greg KH wrote:
> On Wed, May 14, 2008 at 12:09:33AM -0700, Andrew Morton wrote:
> > 
> > =============================================
> > [ INFO: possible recursive locking detected ]
> > 2.6.26-rc2-mm1 #15
> > ---------------------------------------------
> > modprobe/942 is trying to acquire lock:
> >  (&cls->mutex){--..}, at: [<ffffffff811b431e>] device_add+0x43d/0x57a
> > 
> > but task is already holding lock:
> >  (&cls->mutex){--..}, at: [<ffffffff811b6787>] class_interface_register+0x48/0xbd
> > 
> > other info that might help us debug this:
> > 1 lock held by modprobe/942:
> >  #0:  (&cls->mutex){--..}, at: [<ffffffff811b6787>] class_interface_register+0x48/0xbd
> > 
> > stack backtrace:
> > Pid: 942, comm: modprobe Not tainted 2.6.26-rc2-mm1 #15
> > 
> > Call Trace:
> >  [<ffffffff81056be1>] __lock_acquire+0x90d/0xc50
> >  [<ffffffff8100c50f>] ? restore_args+0x0/0x30
> >  [<ffffffff811b431e>] ? device_add+0x43d/0x57a
> >  [<ffffffff81057276>] lock_acquire+0x91/0xb7
> >  [<ffffffff811b431e>] ? device_add+0x43d/0x57a
> >  [<ffffffff812b23ab>] mutex_lock_nested+0xf2/0x278
> >  [<ffffffff811b431e>] ? device_add+0x43d/0x57a
> >  [<ffffffff812b3acd>] ? _spin_unlock+0x23/0x28
> >  [<ffffffff811b431e>] device_add+0x43d/0x57a
> >  [<ffffffff811b4471>] device_register+0x16/0x1b
> >  [<ffffffff811b4555>] device_create+0xdf/0x112
> >  [<ffffffff81055fdc>] ? trace_hardirqs_on+0xd/0xf
> >  [<ffffffff81055fdc>] ? trace_hardirqs_on+0xd/0xf
> >  [<ffffffff812b1fe1>] ? mutex_unlock+0x9/0xb
> >  [<ffffffff811b79b7>] ? kobj_map+0x113/0x124
> >  [<ffffffff810ac0f5>] ? exact_lock+0x0/0x14
> >  [<ffffffff810abd09>] ? exact_match+0x0/0x9
> >  [<ffffffffa00ec448>] :sg:sg_add+0x2a3/0x3bd
> >  [<ffffffff811b67b6>] class_interface_register+0x77/0xbd
> >  [<ffffffffa005d869>] :scsi_mod:scsi_register_interface+0x11/0x13
> >  [<ffffffffa00d50a3>] :sg:init_sg+0xa3/0x155
> >  [<ffffffff8105ea8f>] sys_init_module+0x1823/0x197a
> >  [<ffffffff810c45bc>] ? seq_release+0x0/0x56
> >  [<ffffffff8100bebb>] system_call_after_swapgs+0x7b/0x80
> 
> I'm guessing that this is due to David's "change the semaphore to a
> mutex" patch that you have in your tree, but I refused to take as I was
> worried about just this issue :)

Hm, I thought I saw the same patch from Arjan months ago ... anyway.

class_interface_register() (at least in -rc2) holds &parent->sem (ie the
result of calling class_get() on the class) around calling ->add_dev.
The class in this case is the sg_sysfs_class, so it's calling sg_add().
sg_add() calls device_create() which calls device_add() which acquires
the &dev->class->sem.  The class in this case would _also_ seem to be
sg_sysfs_class.

So why doesn't this deadlock today?

-- 
Intel are signing my paycheques ... these opinions are still mine
"Bill, look, we understand that you're interested in selling us this
operating system, but compare it to ours.  We can't possibly take such
a retrograde step."