Re: 2.6.31rc5 RAID10 lockdep report - sysfs_nofity_dirent locking issue

[Greg KH <gregkh@suse.de>]

On Wed, Aug 05, 2009 at 07:13:08AM +1000, NeilBrown wrote:
> On Wed, August 5, 2009 12:14 am, Dave Jones wrote:
> > Report from a user we received today ..
> > (https://bugzilla.redhat.com/show_bug.cgi?id=515471)
> >
> > 	Dave
> >
> >
> > =================================
> > [ INFO: inconsistent lock state ]
> > 2.6.31-0.118.rc5.fc12.x86_64 #1
> > ---------------------------------
> > inconsistent {SOFTIRQ-ON-W} -> {IN-SOFTIRQ-W} usage.
> > md126_resync/475 [HC0[0]:SC1[1]:HE1:SE0] takes:
> > (sysfs_open_dirent_lock){+.?...}, at: [<ffffffff811a7d08>]
> > sysfs_notify_dirent+0x2c/0x75
> 
> So the problem is that sysfs_open_dirent_lock is sometimes called
> with SOFTIRQs enabled, and sometimes called in a SOFTIRQ and this
> can lead to a deadlock.
> 
> The call from in a softirq is in sysfs_notify_dirent which is a very
> small atomic function which I would certainly like to be able to
> call from any context.
> 
> So the fix would be to use spin_lock_irq on the sysfs_open_dirent_lock ??
> Or should it be spin_lock_bh ??
> 
> Attached is a suggestion for a patch.
> 
> Greg: does that look right?  If so I'll add a changelog entry and
> submit it properly (after at least a compile test...)

Yes, it looks correct, if it passes your tests :)

thanks,

greg k-h