module refcnt underflow

module refcnt underflow

[Lucas De Marchi]

Hi,

While checking somethings for a next release of kmod I noticed this
strange thing:

cat /sys/module/acpi_cpufreq/refcnt
18446744073709551614

We outputs this as "-1" in lsmod because we treat the value as signed.
I was just going to change it to unsigned to make it compatible with
what module-init-tools does but... For me it looks like a bug in the
kernel since in this file we just have the output of
kernel/module.c:module_refcount()

I have this behavior on 3.9.2 but judging by the missing commits in
kernel/module.c, this would also happen in latest head. I've never
seen such a value so I guess it might be difficult to reproduce it
again.

Any hints on what could cause decs to be greater than incs in module_refcount()?


--

Lucas De Marchi

Re: module refcnt underflow

[Rusty Russell]

Lucas De Marchi <lucas.de.marchi@gmail.com> writes:
> Hi,
>
> While checking somethings for a next release of kmod I noticed this
> strange thing:
>
> cat /sys/module/acpi_cpufreq/refcnt
> 18446744073709551614
>
> We outputs this as "-1" in lsmod because we treat the value as signed.
> I was just going to change it to unsigned to make it compatible with
> what module-init-tools does but... For me it looks like a bug in the
> kernel since in this file we just have the output of
> kernel/module.c:module_refcount()
>
> I have this behavior on 3.9.2 but judging by the missing commits in
> kernel/module.c, this would also happen in latest head. I've never
> seen such a value so I guess it might be difficult to reproduce it
> again.
>
> Any hints on what could cause decs to be greater than incs in module_refcount()?

That points to a bug in the cpufreq code.

At a glance, there's nothing obvious, but there are quite a few
reference counts in that code.

Informed people cc'd...

Thanks,
Rusty.