Re: [PATCH 1/3] xen/mce: Add mcelog support for Xen platform (RFC)

[Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>]

On Wed, May 30, 2012 at 03:09:12PM +0000, Liu, Jinsong wrote:
> > 
> > Still no go, this is current linus with your patch applied. I'll look
> > into it 
> > later when there's time.
> 
> The root cause is,
> 1). at cpu/mcheck/mce.c, device_initcall_sync(mcheck_init_device) is *after* all device_initcall();
> 2). at cpu/mcheck/mce_amd.c, device_initcall(threshold_init_device) will
>     threshold_init_device 
>     --> threshold_create_device 
>     --> threshold_create_bank
>     --> kobject_create_and_add(name, &dev->kobj);
>         // at this point, struct device *dev = per_cpu(mce_device, cpu), which is a NULL pointer.
>         // mce_device is initialized at mcheck_init_device --> mce_device_create
> 3). so kernel panic
> 
> So our RFC patch would affect amd mce logic.
> 
> ===========================
> 
> I have a thought about symlink approach, but seems it would bring more issues, e.g.
> 1). it need change more native mce code, like remove /dev/mcelog which created at native mce (under xen platform), or
> 2). it still need to change device_initcall(mcheck_init_device) to device_initcall_sync(mcheck_init_device), if it want to implicitly block native /dev/mcelog --> but that would panic amd mce logic.
> 
> IMO currently there are 2 options:
> 1). use the original approach (implicitly redirect /dev/mcelog to xen_mce_chrdev_device) --> what point of this approach do you think unreasonable? It just remove a 'static' from native mce code!
> 2). use another /dev/xen-mcelog interface, with another misc minor '226'

The 2) is no good.

3) What about moving the corresponding other users (so threshold_init_device), 
to be at late_initcall and the mce to be at late_initcall_sync

4) Or make the driver you are making start at fs_initcall ?