Re: [PATCH 2/3] printk: kunit: support offstack cpumask

[Thomas Weißschuh]

On Wed, Jul 09, 2025 at 01:36:14PM +0200, Petr Mladek wrote:
> On Tue 2025-07-08 16:48:47, Arnd Bergmann wrote:
> > On Tue, Jul 8, 2025, at 16:24, Petr Mladek wrote:
> > > On Wed 2025-07-02 13:28:35, Nathan Chancellor wrote:
> > >> On Wed, Jul 02, 2025 at 11:51:56AM +0200, Petr Mladek wrote:
> > >
> > > Thanks a lot for the nice report.
> > >
> > > The problem is how cpumask_var_t is defined in include/linux/cpumask_types.h:
> > >
> > > #ifdef CONFIG_CPUMASK_OFFSTACK
> > > typedef struct cpumask *cpumask_var_t;
> > > #else
> > > typedef struct cpumask cpumask_var_t[1];
> > > #endif /* CONFIG_CPUMASK_OFFSTACK */
> > >
> > > And KUNIT_DEFINE_ACTION_WRAPPER() expect that the 3rd parameter
> > > is a pointer.
> > >
> > > I am going to solve this by adding a wrapper over free_cpumask_var()
> > > which would work with a pointer to cpumask_var_t.
> > 
> > I'm not familiar enough with the cleanup mechanism of kunit,
> > but can't you just move the mask allocation outside of
> > test_readerwriter()?
> 
> The only solution would be global variable.

When the cpumask is allocated on the stack, free_cpumask_var() is a no-op.
So while the stack address would be leaked to another thread,
it should be fine as nothing is ever done with it.
For more clarity it could also be gated explicitly:

if (IS_ENABLED(CONFIG_CPUMASK_OFFSTACK)) {
	err = kunit_add_action_or_reset(test, prbtest_cpumask_cleanup, test_cpus);
	KUNIT_ASSERT_EQ(test, err, 0);
}

<snip>