Re: kref refcounting breakage in mainline

[Greg KH <greg@kroah.com>]

On Thu, Mar 15, 2007 at 08:53:07AM +0100, Mike Galbraith wrote:
> On Wed, 2007-03-14 at 22:27 -0700, Greg KH wrote:
> 
> > But can you try this version instead?
> 
> It exploded in strcmp().  Unfortunately, the full oops didn't make it to
> either console or serial console.
> 
> [   30.783048] ipmi message handler version 39.1
> [   30.787632] ipmi device interface
> [   30.791166] IPMI System Interface driver.
> [   30.816961] BUG: unable to handle kernel NULL pointer dereference at virtual address 00000000
> [   30.832700]  printing eip:
> [   30.842383] c02d4098
> [   30.851458] *pde = 00000000
> [   30.861089] Oops: 0000 [#1]
> [   30.870585] PREEMPT SMP 
> [   30.879724] Modules linked in:
> [   30.889288] CPU:    1
> [   30.889290] EIP:    0060:[<c02d4098>]    Not tainted VLI
> [   30.889292] EFLAGS: 00010282   (2.6.21-rc3-smp #81)

Gah, that just happened to me too, sorry for not booting with that
change before sending it to you.

> I did this...
> 
> --- kernel/module.c.org	2007-03-15 07:20:15.000000000 +0100
> +++ kernel/module.c	2007-03-15 08:13:36.000000000 +0100
> @@ -2419,15 +2419,19 @@ void module_remove_driver(struct device_
>  
>  	if (drv->owner && drv->owner->mkobj.drivers_dir)
>  		mk = &drv->owner->mkobj;
> -	else {
> +	else if (drv->mod_name) {
>  		/* Lookup built-in module entry in /sys/modules */
>  		mkobj = kset_find_obj(&module_subsys.kset, drv->mod_name);
>  		if (!mkobj)
> -			return;
> +			goto out_free;
>  		mk = container_of(mkobj, struct module_kobject, kobj);
> +	} else {
> +		WARN_ON(1);
> +		goto out_free;
>  	}
>  	sysfs_remove_link(mk->drivers_dir, driver_name);
>  	kobject_put(mkobj);
> +out_free:
>  	kfree(driver_name);
>  }
>  EXPORT_SYMBOL(module_remove_driver);
> 
> ...and it booted.

That's good.  But why don't we have a module name for this driver?

And if we don't have a module name, why would there be a symlink to
remove?  That's what is keeping your module from unloading, right?

> [   24.670410] ipmi message handler version 39.1
> [   24.675000] ipmi device interface
> [   24.678542] IPMI System Interface driver.
> [   24.703956] BUG: at kernel/module.c:2429 module_remove_driver()
> [   24.716837]  [<c0105086>] show_trace_log_lvl+0x1a/0x30
> [   24.728909]  [<c01057b2>] show_trace+0x12/0x14
> [   24.740239]  [<c0105856>] dump_stack+0x16/0x18
> [   24.751469]  [<c01441ce>] module_remove_driver+0xa5/0xa7
> [   24.763584]  [<c035a014>] bus_remove_driver+0x6d/0x82
> [   24.775390]  [<c035a978>] driver_unregister+0xb/0x18
> [   24.787019]  [<c034c8c2>] init_ipmi_si+0x7a9/0x7c1
> [   24.798450]  [<c06705bc>] init+0x144/0x26c
> [   24.809129]  [<c0104cfb>] kernel_thread_helper+0x7/0x1c
> [   24.820916]  =======================
> [   24.830926] ipmi_si: Unable to find any System Interface(s)
> [   24.842952] IPMI Watchdog: driver initialized
>   24.853749] Copyright (C) 2004 MontaVista Software - IPMI Powerdown via sys_reboot.

With the above change, it all works correctly?

If the ipmi driver is loaded, what does the /sys/module/MODULE_NAME/
tree look like (replacing MODULE_NAME with whatever the module name
really is, sorry, I don't know)?

thanks,

greg k-h