Re: [OOPS] sysfs_hash_and_remove (was Re: What protection ....)

[James Bottomley <James.Bottomley@HansenPartnership.com>]

On Thu, 2005-11-24 at 15:34 +0100, Ingo Molnar wrote:
> * Maneesh Soni <maneesh@in.ibm.com> wrote:
> 
> > So, IMO, it is necessary to explicitly remove links before 
> > unregistering the kobject in case of bidirectional cross symlinks.
> > 
> > The patch from James, is working, because it is not creating the cross 
> > symlink itself.
> 
> so, what is your suggestion, what should be done to fix the problem? The 
> patch below:
> isnt fit for upstream inclusion :-)

Well, the patch was just intended to confirm the problem diagnosis.

The solution Maneesh appears to be advocating to the issue is imposing
del ordering, the issue being that device_del is the call that actually
removes the directories and symlinks, so all callers have to make sure
they've called class_device_del for every class on the device before
calling device_del.  Since this trigger point happens regardless of
references, we can't expect references to get us out of this one, so
we'll have to audit the failing code manually.  I did find and fix one
of these issues in SCSI, but there may be more lurking around ...

James