Re: [RFC] kobject: add kobject_init_ng and kobject_init_and_add functions

[Greg KH <greg@kroah.com>]

On Fri, Nov 30, 2007 at 03:25:52PM -0500, Alan Stern wrote:
> On Fri, 30 Nov 2007, Greg KH wrote:
> 
> > +/**
> > + * kobject_init_and_add - initialize a kobject structure and add it to the kobject hierarchy
> > + * @kobj: pointer to the kobject to initialize
> > + * @ktype: pointer to the ktype for this kobject.
> > + * @parent: pointer to the parent of this kobject.
> > + * @fmt: the name of the kobject.
> > + *
> > + * This function will properly initialize a kobject and then call
> > + * kobject_add().
> > + *
> > + * If the function returns an error, the memory allocated by the kobject
> > + * can be safely freed, no other functions need to be called.
> > + */
> > +int kobject_init_and_add(struct kobject *kobj, struct kobj_type *ktype,
> > +			 struct kobject *parent, const char *fmt, ...)
> > +{
> > +	va_list args;
> > +	int retval;
> > +
> > +	va_start(args, fmt);
> > +	retval = kobject_init_varg(kobj, ktype, parent, fmt, args);
> > +	va_end(args);
> > +	if (retval)
> > +		return retval;
> > +
> > +	retval = kobject_add(kobj);
> > +	if (retval)
> > +		kobject_put(kobj);
> 
> No, no!
> 
> You have recreated the problem we have been discussing during the last
> couple of days.  If the kobject_init_varg() routine gets an error then
> the kobject will need to be deallocated manually.  If the kobject_add()
> routine gets an error then the cleanup invoked by kobject_put() will do
> the deallocation automatically.
> 
> But the caller can't tell in which subroutine an error occurred, so it
> won't know what to do when kobject_init_and_add() returns an error.

Oh crap.  You're totally right.  I suck.

> The only way to resolve this problem is to have the _init routine 
> consume no resources and never fail.  That way the only possible 
> failure mode would be if the _add routine doesn't work, in which case 
> either a kfree() or a kobject_put() would be acceptable.
> 
> In particular, this implies that the name should be set as part of the 
> _add() call, not as part of _init().  This is more in line with the way 
> the code tends to use kobjects anyhow.  Unless people want to name 
> unregistered kobjects -- does this ever happen?  And it if does, can 
> these kobjects simply be replaced by krefs?

No, the only non-registered kobjects in the tree right now are never
named.  So this should be safe.

> My suggestion: Have kobject_init_ng() accept a ktype pointer but not a 
> parent or name.  Instead, make kobject_add_ng() take the parent and 
> name (possibly a kset also).  Then when kobject_init_and_add() 
> encounters an error, it shouldn't do a _put() -- the caller can either 
> do the _put() or just do a kfree().

Why not the parent for init()?  Isn't it always known at that time?
I'll dig to be sure.

Ok, second round of patches coming up...

thanks,

greg k-h