On Sun, Jul 20, 2008 at 12:27:26AM +0200, Vegard Nossum wrote:

On Sun, Jul 20, 2008 at 12:17 AM, Greg KH [off-list ref] wrote:

quoted

On Sat, Jul 19, 2008 at 02:59:12PM +0200, Vegard Nossum wrote:

quoted

On Sat, Jul 19, 2008 at 11:55 AM, Vegard Nossum [off-list ref] wrote:

quoted

What I don't get here is how SLUB can be used this early in the boot
process. Notice that this is still miles away from the

   SLUB: Genslabs=12, HWalign=128, Order=0-3, MinObjects=0, CPUs=1, Nodes=1

line, which comes much later. And that kobject_init() _is_ calling
kzalloc() via verify_dynamic_kobject_allocation(). Isn't this an
error?

(Unfortunately, my "git log" doesn't turn up any recent changes for
any of the affected code paths here.)

Ehe... and this is the reason why: The code was added by this patch:

commit 0e3638d1e04040121af00195f7e4628078246489
Author: Dave Hansen [off-list ref]
Date:   Thu Mar 16 17:30:16 2006 -0800

    warn when statically-allocated kobjects are used

..which only exists in -next. Is that just a truly ancient patch, or
did somebody forget to adjust their clock?

It is truely a very old patch, that only lives in my tree, and currently
isn't planned to go to Linus any year soon.

But it has a very long history of living in the -mm tree, and finding
real bugs, it's just not "safe" enough to go to Linus's tree.  Unless
you think it is?

Hm. In this case, the patch is not even reporting a problem, it is in
fact in error itself.

The problem is that it calls kzalloc() before the slab caches have
been set up. (Yes, it's a wonder that nothing crashed.) I can only
suggest the addendum

    if (!slab_is_available())
        return;

That's odd, what changed to cause this to become a problem?  Is a
kobject somehow now being created too early in the boot process than it
was in the past?  This code hasn't changed in a very long time, so I'm
loath to blame this code, odds are the root cause is somewhere else...

thanks,

greg k-h