Re: Fixing the main programmer thinko with the device model

[Greg KH <greg@kroah.com>]

On Wed, Mar 26, 2008 at 06:07:49AM +0100, Kay Sievers wrote:
> On Tue, 2008-03-25 at 21:16 -0700, Greg KH wrote:
> > On Tue, Mar 25, 2008 at 10:57:32AM +0100, Andi Kleen wrote:
> > > James Bottomley <James.Bottomley@HansenPartnership.com> writes:
> > > > 
> > > > That's true, but irrelevant (and also soon to be untrue if we get rid of
> > > > the scsi_device class as you and Kay keep requesting).  The two calls
> > > > release references on the actual embedded generic device, it's nothing
> > > > to do with entangled lifetime rules.
> > > 
> > > <heretic thought>Has anybody ever considered just doing away with 
> > > the problematic and bug prone and tricky reference counts for kobjects 
> > > and switch to a simple garbage collector for them?
> > 
> > Sure, I have no objection to that.  It's just that the reference count
> > "issue" really doesn't seem to be one on sanely designed busses :)
> 
> Hmm, what is a "???simple garbage collector" here? How could one determine
> "reachability" of objects, means: at what point of time do objects
> actually become "garbage"? How could one trace in our current kernel
> code who still accesses an object, without doing refcounts?

I think in the end, it would be the same thing, so we should stick with
our current code, as that's exactly what the current kobject model now
implements :)

thanks,

greg k-h