RE: [RFC] kref, a tiny, sane, reference count object

["Peter Kjellerstedt" <peter.kjellerstedt@axis.com>]

> -----Original Message-----
> From: linux-kernel-owner@vger.kernel.org 
> [mailto:linux-kernel-owner@vger.kernel.org] On Behalf Of Greg KH
> Sent: Saturday, March 13, 2004 09:20
> To: linux-kernel@vger.kernel.org
> Subject: [RFC] kref, a tiny, sane, reference count object
> 
> In thinking about people's complaints about the current 
> kobject interface, a lot of people don't like the 
> "complexity" of what is necessary to use a kobject.
> If all you want is something to handle reference 
> counting properly, a kobject can seem a bit "large".
> 
> For all of those people, this patch is for you.  
> Introducing struct kref.  A tiny (only 8 bytes on a 
> 32bit platform) that will properly handle reference 
> counting any structure you want to use it for.  Note 
> that you will have to be careful around the cleanup 
> period (but that can be easily handled by the user 
> with regards to not trying to grab a "new" reference
> if you don't already have one, once the object is 
> gone, just like kobjects and sysfs today work.)
> 
> I've implemented kobjects using a kref to handle the 
> reference counting portion, but will leave that patch
> and change for 2.7, as it will add 4 more bytes (on a 
> 32bit platform) to every kobject, and that wouldn't 
> be nice this early in the 2.6 series.  For now, krefs 
> can stand on their own.
> 
> I've already found loads of places in the kernel that 
> can use this structure to clean up their logic, and 
> will probably be converting a number of them over time 
> to use them.  But no, Al, I will not say this can be 
> used to replace the atomic_t count you have in inodes, 
> as that count is horribly abused in ways I never really 
> wanted to know about (negative counts mean something 
> "special"?  eeeeeek....)
> 
> Anyway, here's a patch against 2.6.4 that adds krefs
> to the kernel.  I'll follow up with a patch that 
> converts the usb-serial core from using kobjects to 
> using krefs instead.
> 
> Comments are appreciated and welcomed.
> 
> thanks,
> 
> greg k-h

Looks simple enough.  But I have a small question. 
In kref_get() and kref_cleanup(), kref is verified 
not to be NULL before being used.  However, this is 
not done in kref_put().  An oversight, or as intended?

//Peter