Re: [PATCH] sysfs: handle duplicate removal attempts in sysfs_remove_group()

["Rafael J. Wysocki" <rjw@rjwysocki.net>]

On Monday, November 25, 2013 10:29:00 AM James Bottomley wrote:
> On Fri, 2013-11-22 at 11:02 -0500, Tejun Heo wrote:
> > Hello,
> > 
> > On Fri, Nov 22, 2013 at 08:43:55AM -0700, Bjorn Helgaas wrote:
> > > > So, we do have cases where the parent is removed before the child.  I
> > > > suppose the parent pci bridge is removed already?  AFAICS this
> > > > shouldn't break anything but people did seem to expect the removals to
> > > > be ordered from child to parent.  Bjorn, is this something you expect
> > > > to happened?
> > > 
> > > I do not expect a PCI bridge to be removed before the devices below
> > > it.  We should be removing all the children before removing the parent
> > > bridge.
> > > 
> > > But is this related to PCI?  I don't see the connection yet.  I tried
> > 
> > I'm not sure.  It was from thunderbolt and nobody is reporting it on
> > other interconnects, so it could be.
> > 
> > > to look into this a bit (my notes are at
> > > https://bugzilla.kernel.org/show_bug.cgi?id=65281), but I haven't
> > > figured out the big-picture problem yet.
> > > 
> > > I don't have warm fuzzies that adding a "have we already removed this"
> > > check is the best resolution, but maybe that's just because I don't
> > > understand the problem.
> > 
> > Yeah, the whole thing is sorta pointless.  Just issuing removal and
> > continuing on should do, IMHO.
> 
> I'd go for that as well.  We have huge problems with the _del calls
> because visibility is strict hierarchy and it's not always easy to work
> out who's underneath us.
> 
> It's going to be really annoying when refcounting works perfectly for
> objects, so you can just do puts in any order, but you have to have
> _del() called to remove subordinate objects before their parent.

Well, that would be fine and dandy, but device_del() calls bus_remove_device()
which in turn runs device_release_driver() and the order in which *these*
things are done actually matters in general.

So yes, after we have released all of the drivers in question, we can do _del()
right before the final _put() in any order just fine.

Thanks,
Rafael