Hello,

On Thu, 2 May 2013, Peter Zijlstra wrote:

On Wed, May 01, 2013 at 09:22:08PM +0300, Julian Anastasov wrote:

quoted

quoted

+extern int __cond_resched_rcu(void);
+
+#define cond_resched_rcu() ({			\
+	__might_sleep(__FILE__, __LINE__, 0);	\

	I see your goal. But digging into __might_sleep()
I see that rcu_sleep_check() will scream for the non-preempt
case because we are under rcu_read_lock.

#ifdef CONFIG_PREEMPT_RCU
#define PREEMPT_RCU_OFFSET 0
#else
#define PREEMPT_RCU_OFFSET 1
#endif

#define cond_resched_rcu() ({	\
	__might_sleep(__FILE__, __LINE__, PREEMPT_RCU_OFFSET);	\
	__cond_resched_rcu();	\
})

Should work I think..

	Looks like CONFIG_DEBUG_ATOMIC_SLEEP selects
CONFIG_PREEMPT_COUNT, so PREEMPT_RCU_OFFSET should be
1 in all cases because preempt_disable() adds 1, while
for CONFIG_PREEMPT_RCU case rcu_preempt_depth() should
return 1:

#ifdef CONFIG_PREEMPT_RCU
#define PREEMPT_RCU_OFFSET 1
#else
#define PREEMPT_RCU_OFFSET PREEMPT_CHECK_OFFSET
#endif

	Now the remaining part is to fix rcu_sleep_check() for
the non-preempt case. As there are no nesting depths in this
case, I don't see a solution so far. We can provide
some argument to rcu_preempt_sleep_check to compare
depth with preempt_count() but currently I don't know
how to differentiate cond_resched_lock() from cond_resched_rcu()
when calling __might_sleep, in both cases we provide
PREEMPT_OFFSET. May be some trick is needed here without
adding new arg to __might_sleep, so that we can properly
check for rcu_lock_map.

Regards

--
Julian Anastasov [off-list ref]