Re: [PATCH] EDAC: core EDAC support code

[Greg KH <greg@kroah.com>]

On Tue, Mar 07, 2006 at 08:03:41PM -0500, Dmitry Torokhov wrote:
> On Tuesday 07 March 2006 12:04, Greg KH wrote:
> > On Tue, Mar 07, 2006 at 08:47:44AM -0800, Dave Peterson wrote:
> > > Ok, how does this sound:
> > > 
> > >     - Modify EDAC so it uses kmalloc() to create the kobject.
> > >     - Eliminate edac_memctrl_master_release().  Instead, use kfree() as
> > >       the release method for the kobject.  Here, it's important to use a
> > >       function -outside- of EDAC as the release method since the core
> > >       EDAC module may have been unloaded by the time the release method
> > >       is called.
> > 
> > No, if this happens then you are using the kobject incorrectly.  How
> > could it be held if your module is unloaded?  Don't you have the module
> > reference counting logic correct?
> > 
> 
> It is pretty hard to implement kobject handling correctly. Consider the
> following:
> 
> 	rmmod device_driver < /sys/devices/pci0000:00/...../power/state
> 
> for a driver that creates/destroys device objects.

I agree, that's one reason I really hate the "default" attributes :(

To do this "right" we need to make the attributes dynamically created
and the owner set to the proper module.  I did that for the module core
code and it's on my todo list for the driver core too.

> Opening 'state' attribute will pin device structure into memory but will
> not increase _your_ module's refcount. It is nice if you have a subsystem
> core split from drivers code - then you can keep core module reference
> until device objects are gone and allow individual drivers be unloaded
> freely. But for single-module system it is pretty hard, that's why
> platform devices are popular.

They are popular for when you don't have a "bus", and rightfully so.

thanks,

greg k-h