kobject must be initialized before calling kobject_init()?!

kobject must be initialized before calling kobject_init()?!

[Roland Dreier]

So I was perusing the code in lib/kobject.c, and I saw this:

	void kobject_init(struct kobject *kobj, struct kobj_type *ktype)
	{
		// [a couple of of parameter checks...]
		if (kobj->state_initialized) {
			/* do not error out as sometimes we can recover */
			printk(KERN_ERR "kobject (%p): tried to init an initialized "
			       "object, something is seriously wrong.\n", kobj);
			dump_stack();
		}

in other words the first thing you do is check a member of the
structure you're supposed to be initializing -- if someone just
kmalloc()s a struct kobject and passes it into this function, there's
a good chance that state_initialized won't be zero.  In fact, with
slab debugging on, it's guaranteed to be poisoned with a non-zero
value.

So are users supposed to be zeroing out struct kobjects before
initializing them?  If so, this should probably be documented.
And also the code in kobject_init() that sets other fields to be zero
can also be removed:

	kobj->state_in_sysfs = 0;
	kobj->state_add_uevent_sent = 0;
	kobj->state_remove_uevent_sent = 0;

Let me know if the intention is that everyone should zero out their
kobjects when they initialize them, and I'll send a patch updating the
documentation.

There's another similar booby-trap in kobject_init_and_add(), which
ends up calling into kobject_set_name_vargs(), which does

	/* Free the old name, if necessary. */
	kfree(kobj->name);

which is generally a bad thing if kobj->name has not been initialized.

Thanks,
  Roland

Re: kobject must be initialized before calling kobject_init()?!

[Greg KH]

On Fri, Feb 08, 2008 at 08:24:25PM -0800, Roland Dreier wrote:
> So I was perusing the code in lib/kobject.c, and I saw this:
> 
> 	void kobject_init(struct kobject *kobj, struct kobj_type *ktype)
> 	{
> 		// [a couple of of parameter checks...]
> 		if (kobj->state_initialized) {
> 			/* do not error out as sometimes we can recover */
> 			printk(KERN_ERR "kobject (%p): tried to init an initialized "
> 			       "object, something is seriously wrong.\n", kobj);
> 			dump_stack();
> 		}
> 
> in other words the first thing you do is check a member of the
> structure you're supposed to be initializing -- if someone just
> kmalloc()s a struct kobject and passes it into this function, there's
> a good chance that state_initialized won't be zero.  In fact, with
> slab debugging on, it's guaranteed to be poisoned with a non-zero
> value.
> 
> So are users supposed to be zeroing out struct kobjects before
> initializing them?

Yes.  See also the check about the non-zero reference count when
initializing the kref, that too will trip.

> If so, this should probably be documented.

You're right.  Patches gladly accepted :)

thanks,

greg k-h