Re: 2.6.27.14: BUG: lock held when returning to user space!

[Jan Kara <jack@suse.cz>]

> On Wed, 2009-03-04 at 17:58 +0100, Jan Kara wrote:
> > > On Wed, 2009-02-25 at 15:21 +0100, Frank van Maarseveen wrote:
> > > > An lvextend -L+16G command for a logical volume while being mounted rw
> > > > as ext3 triggered the following on 2.6.27.14:
> > > > 
> > > > ================================================
> > > > [ BUG: lock held when returning to user space! ]
> > > > ------------------------------------------------
> > > > lvextend/29191 is leaving the kernel with locks still held!
> > > > 2 locks held by lvextend/29191:
> > > >  #0: (&type->s_umount_key #15){....}, at: [<c019edfb>] get_super+0x6b/0xb0
> > > >  #1: (&journal->j_barrier){....}, at: [<c01f9af3>] journal_lock_updates+0xc3/0xd0
> > > 
> > > Do recent kernels still say this?
> >
> >   I'd say so. We really hold j_barrier mutex when leaving the kernel
> > after FIFREEZE ioctl (the call path goes as freeze_bdev -> ext3_freeze
> > -> journal_lock_updates) until FITHAW is called. As far as I know it was
> > designed this way...
> >   It would be a pity to completely exclude j_barrier mutex from lockdep
> > control so would it be possible to mark the mutex (or even that
> > particular acquisition of the mutex) so that lockdep does not warn when
> > we return with it to userspace?
> 
> Linus specificly stated that we are not to keep locks held when
> returning to userspace:
> 
>   http://lkml.org/lkml/2007/10/27/135
> 
> So sure, we could annotate this, but no I won't until you can convince
> both me and Linus that its a sane thing to do.
  I didn't write the code ;). I was just informing how the things are. I agree
it's not a nice thing to do. Hmm, looking at the code, it uses a special
semaphore bd_mount_sem while mutex seems to be enough. Digging in git...
ah, exactly so that lockdep does not complain about it. Ugh. Also
the comment before freeze_bdev() says something about taking s_umount while
it does not take it. That code really seems to need some cleaning...

> The problems include: how can you be sure its the same task calling the
> completing ioctl?
  Yeah, I understand the problems.

									Honza
-- 
Jan Kara <jack@suse.cz>
SuSE CR Labs