Re: 2.6.17-rc5-mm2

[Andrew Morton <akpm@osdl.org>]

On Thu, 01 Jun 2006 19:27:41 +0200
Arjan van de Ven <arjan@linux.intel.com> wrote:

> 
> > > http://www.stardust.webpages.pl/files/mm/2.6.17-rc5-mm2/bug_2.jpg
> > 
> > So it's claiming that we're taking multiple i_mutexes.
> > 
> > I can't immediately see where we took the outermost i_mutex there.
> 
> inlining caused one level to be removed from the backtrace
> one level is in fs_remove_file, the sub level is usbfs_unlink (called
> from fs_remove_file)

OK.

I'll duck this patch for now, pending a tested-n-changelogged one, please.

> >   Nor is
> > it immediately obvious why this is considered to be deadlockable?
> 
> what is missing is that we tell lockdep that there is a parent-child
> relationship between those two i_mutexes, so that it knows that 1)
> they're separate and 2) that the lock take order is parent->child
> 
> 
> > (lockdep tells us that a mutex was taken at "mutex_lock+0x8/0xa", which is
> > fairly useless.  We need to report who the caller of mutex_lock() was).
> 
> yeah this has been bugging me as well; either via a wrapper around
> mutex_lock or via the gcc option to backwalk the stack (but that only
> works with frame pointers enabled.. sigh)

Actually, __builtin_return_address(0) works OK with -fomit-frame-pointer,
and that's all we need here.