Re: x86/mce: machine check warning during poweroff

[Linus Torvalds <torvalds@linux-foundation.org>]

On Fri, Jan 13, 2012 at 12:22 PM, Srivatsa S. Bhat
<srivatsa.bhat@linux.vnet.ibm.com> wrote:
>
> Fundamentally, this warning is triggered during CPU Offline, which is done
> during poweroff, suspend, hibernate etc. IOW, even a simple
> # echo 0 > /sys/devices/system/cpu/cpuX/online will trigger it.

There is definitely something wrong with CPU hotplug and MCE.

I seem to be able to trigger not only warnings, but some oopses, by doing:

 - enable list debugging, slab debugging, and kobject debugging in the
kernel (I've got some other things enabled too, but I think those are
the main ones)

 - do

     echo 0 > /sys/devices/system/cpu/cpuX/online

   this gets a few warnings

 - then do

     echo 1 > /sys/devices/system/cpu/cpuX/online

where bringing it up again will crash the machine entirely.

The oops scrolled off the screen an ddidn't get caught anywhere, but
the call trace seems to be (warning: hand-entered, so some of this may
be bogus):

Oops in:
  kobject_get+0x10/0x40

Code:
  55 48 89 f8 48 89 e5 48 83 ec 10 48 85 ff 74 0b <8b> 57 38 85 d2 74 06 f0 ff

Call trace:
  get_device
  klist_device_get
  klist_add_tail
  bus_add_device
  device_add
  device_register
  mce_device_create
  notifier_call_chain
  __raw_notifier_call_chain
  __cpu_notify
  _cpu_up
  store_online
  dev_attr_change
  sysfs_write_file

so it's definitely something bad in MCE device handling, and probably
something to do with reusing a 'struct device' after freeign it, or
after not having completely cleaned it up.

I didn't see if I could spot the problem, but I think this is entirely
reproducible, so hopefully somebody who knows the MCE code can
trivially see this and fix it.

                   Linus