[xfs-masters] Re: Linux 2.6.17-rc2 - notifier chain problem?

[Andrew Morton <akpm@xxxxxxxx>]

Chandra Seetharaman <sekharan@xxxxxxxxxx> wrote:
>
> Thanks for the steps. With that i was able to reproduce the problem and
> i found the bug.
> 
> While i go ahead and generate the patch, i wanted to hear if my
> conclusion is correct.
> 
> The problem is due to the fact that most notifier registrations
> incorrectly use __devinitdata to define the callback structure, as in:
> 
> static struct notifier_block __devinitdata hrtimers_nb = {
>         .notifier_call = hrtimer_cpu_notify,
> };
> 
> devinitdata'd  data is not _expected to be available_ after the
> initialization(unless CONFIG_HOTPLUG is defined).
> 
> I do not know how it was working until now :), anybody has a theory that
> can explain it (or my conclusion is wrong) ?

That sounds right.  There are several __devinitdata notifier_blocks in the
tree - please be sure to check them all.

btw, it'd be pretty trivial to add runtime checking for this sort of thing:

int addr_in_init_section(void *addr)
{
        return addr >= __init_begin && addr < __init_end;
}

(x86-specific)
(need to add __init_end to vmlinux.lds.S)

then we could use that to check various things in various places...