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

[Petr Mladek <pmladek@suse.com>]

On Wed 2025-07-09 14:53:29, Thomas Weißschuh wrote:
> 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);
> }

It is likely a matter of taste but I like this idea. It looks better
than passing an invalid pointer and hope nobody would do anything
with it.

The only problem is that

    if (IS_ENABLED(CONFIG_CPUMASK_OFFSTACK)) {

did not prevented the compiler warning. I guess that the code was still
compiled and later just optimized out.

So, I am going to use #ifdef instead. I create a function:

/*
 * A cast would be needed for the clean up action when the cpumask was on stack.
 * Also it would leak the stack address to the cleanup thread.
 * And alloc_cpu_mask() and free_cpumask_var() would do nothing anyway.
 */
#ifdef CONFIG_CPUMASK_OFFSTACK
KUNIT_DEFINE_ACTION_WRAPPER(prbtest_cpumask_cleanup, free_cpumask_var, cpumask_var_t);

static void prbtest_alloc_cpumask(struct kunit *test, cpumask_var_t *mask)
{
	int err;

	KUNIT_ASSERT_TRUE(test, alloc_cpumask_var(mask, GFP_KERNEL));
	err = kunit_add_action_or_reset(test, prbtest_cpumask_cleanup, *mask);
	KUNIT_ASSERT_EQ(test, err, 0);
}
#else
static inline
void prbtest_alloc_cpumask(struct kunit *test, cpumask_var_t *mask) {}
#endif

which will be called in test_readerwriter().

It seems to work, ..., sigh. I did not expect so many troubles with
a tiny detail.

Best Regards,
Petr