Hello,

On Thu, 2 May 2013, Paul E. McKenney wrote:

On Thu, May 02, 2013 at 06:54:05PM +0300, Julian Anastasov wrote:

quoted

	I tested the following patch in 2 variants,
TINY_RCU and CONFIG_TREE_PREEMPT_RCU. I see the

Could you please also try CONFIG_TREE_RCU?

	Note that I'm testing on some 9-year old
UP system, i.e. 1 CPU. Now I enabled SMP to test CONFIG_TREE_RCU
and the results are same. I think, it should be just like
the TINY_RCU in terms of these debuggings (non-preempt). Extra 
rcu_read_lock gives me "Illegal context switch in RCU read-side
critical section" in addition to the "BUG: sleeping function
called from invalid context" message.

quoted

error if extra rcu_read_lock is added for testing.

	I'm using the PREEMPT_ACTIVE flag to indicate
that we are already under lock. It should work because
__might_sleep is not called with such bit. I also tried to
add new flag in include/linux/hardirq.h but PREEMPT_ACTIVE
depends on the arch, so this alternative looked difficult to
implement.

quoted

+extern int __cond_resched_rcu(void);
+
+#define cond_resched_rcu() ({					\
+	__might_sleep(__FILE__, __LINE__, PREEMPT_ACTIVE |	\
+					  PREEMPT_RCU_OFFSET);	\
+	__cond_resched_rcu();					\
+})
+

quoted

@@ -7062,7 +7076,9 @@ void __might_sleep(const char *file, int line, int preempt_offset)
 {
 	static unsigned long prev_jiffy;	/* ratelimiting */

-	rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
+	/* WARN_ON_ONCE() by default, no rate limit reqd. */
+	rcu_sleep_check(preempt_offset & PREEMPT_ACTIVE);

Color me confused.

quoted

From what I can see, the two values passed in through preempt_offset

are PREEMPT_LOCK_OFFSET and SOFTIRQ_DISABLE_OFFSET.  PREEMPT_ACTIVE
is normally a high-order bit, above PREEMPT_MASK, SOFTIRQ_MASK, and
HARDIRQ_MASK.

PREEMPT_LOCK_OFFSET and SOFTIRQ_DISABLE_OFFSET have only low-order bits,
so I don't see how rcu_sleep_check() is passed anything other than zero.
Am I going blind, or what?

	Only the new cond_resched_rcu() macro provides
PREEMPT_ACTIVE flag to skip the rcu_preempt_sleep_check()
call. The old macros provide locked=0 as you noticed. Does it
answer your question or I'm missing something?

Regards

--
Julian Anastasov [off-list ref]