Re: [RFC 22/22] tcm_loop: Add multi-fabric Linux/SCSI LLD fabricmodule

[Dmitry Torokhov]

On Sun, Sep 05, 2010 at 11:01:01PM -0700, Nicholas A. Bellinger wrote:
> On Sun, 2010-09-05 at 22:38 -0700, Dmitry Torokhov wrote:
> > On Sun, Sep 05, 2010 at 10:21:20PM -0700, Nicholas A. Bellinger wrote:
> > > On Sun, 2010-09-05 at 22:17 -0700, Greg KH wrote:
> > > > On Sun, Sep 05, 2010 at 10:08:16PM -0700, Dmitry Torokhov wrote:
> > > > > On Sun, Sep 05, 2010 at 02:51:19PM -0700, Nicholas A. Bellinger wrote:
> > > > > > On Sun, 2010-09-05 at 13:30 -0700, Dmitry Torokhov wrote:
> > > > > > > On Mon, Aug 30, 2010 at 02:23:23AM -0700, Nicholas A. Bellinger wrote:
> > > > > > > > +
> > > > > > > > +static void tcm_loop_primary_release(struct device *dev)
> > > > > > > > +{
> > > > > > > > +	return;
> > > > > > > > +}
> > > > > > > > +
> > > > > > > > +static struct device tcm_loop_primary = {
> > > > > > > > +	.init_name		= "tcm_loop_0",
> > > > > > > > +	.release		= tcm_loop_primary_release,
> > > > > > > > +};
> > > > > > > > +
> > > > > > > 
> > > > > > > No, you can not have statically allocated devices and dummy release
> > > > > > > functions.
> > > > > > > 
> > > > > > 
> > > > > > Last time I checked this is still what mainline
> > > > > > drivers/scsi/scsi_debug.c is doing for struct device pseudo_primary:
> > > > > > 
> > > > > > static void pseudo_0_release(struct device *dev)
> > > > > > {
> > > > > >         if (SCSI_DEBUG_OPT_NOISE & scsi_debug_opts)
> > > > > >                 printk(KERN_INFO "scsi_debug: pseudo_0_release() called\n");
> > > > > > }
> > > > > > 
> > > > > > static struct device pseudo_primary = {
> > > > > >         .init_name      = "pseudo_0",
> > > > > >         .release        = pseudo_0_release,
> > > > > > };
> > > > > > 
> > > > > > so for these type of things in TCM_Loop I tend to follow what scsi_debug does,
> > > > > >  so I don't exactly see a issue here atm.  Anybody else have comments..?
> > > > > 
> > > > > The kobject embedded in the device structure is refcounted and it's
> > > > > lifetime may be different from lifetime of your module. If somebody
> > > > > happen to hold reference to the driver while your module is being
> > > > > unloaded bad things will happen.
> > > > > 
> > > > > I am sure Greg has something on this topic ready for cut-and-paste.
> > > > 
> > > > Documentation/kobject.txt :)
> > > 
> > > I still don't get it..?  Does this mean that scsi_debug and TCM_Loop are
> > > currently broken wrt to this item or not..?
> > > 
> > 
> > Yes they are broken.
> 
> Hi Dmitry and Greg,
> 
> Would you mind giving a bit more specific idea as to what a patch to fix
> this issue would involve for an emulated struct device usage..?  I would
> be happy to generate patches for scsi_debug and TCM_loop to resolve this
> issue, but I need a bit more information what I should be looking at.
> 

I think what you need is root_device_create() which will allocate a
device structure and assign ->release() that is implemented outside of
your module. This way your module may be unloaded even if there are
outstanding references to tcm_loop_primary device.

The rest of devices you create needs to increment module counter or
otherwise ensure that their methods stay available until they are
cleaned up (->release() is called).

Hope this helps.

-- 
Dmitry
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/