Re: [BUG 2.6.17-git] kmem_cache_create: duplicate cache scsi_cmd_cache

[Al Viro <viro@ftp.linux.org.uk>]

On Fri, May 12, 2006 at 10:55:20PM +0100, Russell King wrote:
> On Fri, May 12, 2006 at 10:43:54PM +0100, Al Viro wrote:
> > On Fri, May 12, 2006 at 09:34:16PM +0100, Russell King wrote:
> > > On Fri, May 12, 2006 at 10:36:57AM -0700, Linus Torvalds wrote:
> > > > Yes. We could just revert that commit, but it seems correct, and I'd 
> > > > really like for somebody to understand _why_ that commit matters at all. I 
> > > > certainly don't see the overlap here..
> > > 
> > > Reverting the commit breaks MMC/SD in a very real way, and the fix
> > > is plainly correct and is actually the only possible fix that can be
> > > applied.
> > 
> > Bullshit.  Could you explain what generic code dereferences ->driverfs_dev
> > after del_gendisk()?  If you see such beast, please tell; _that_ is the
> > real bug.
> 
> Al, I think you're going to eat your own bull on this one.
> 
> You'll find that in the reply I've just sent to Linus - two oopen
> reported by two different people since the mount/umount hotplug
> events got re-merged.
> 
> The problem case is:
> 
> - insert card
> - mount filesystem
> - remove card
> - umount filesystem <bang, oops>
> 
> The generic code is block_uevent(), which is called at umount time
> _after_ the gendisk has been deleted (which happens when the card has
> been removed.)
> 
> Basically, you can't umount a destroyed block device without oopsing.

So block_uevent() is bogus; great, we've located the real bug.  The
real question: what uses these events and what can userland _do_ when
it gets something about a device that had been gone for a month?

Secondary question: who had resurrected that crap?  I distinctly remember
killing it off...