Re: [PATCH -tip] rcu: local_irq_disable() also delimits RCU_SCHED read-site critical sections

[Lai Jiangshan <laijs@cn.fujitsu.com>]

Paul E. McKenney wrote:
> On Tue, Mar 16, 2010 at 07:09:21PM +0800, Lai Jiangshan wrote:
>> It is documented that local_irq_disable() also delimits
>> RCU_SCHED read-site critical sections.
>> See the document of synchronize_sched() or
>> Documentation/RCU/whatisRCU.txt.
>>
>> So we have to test irqs_disabled() in rcu_read_lock_sched_held().
>> Otherwise rcu-lockdep brings incorrect complaint.
> 
> Interesting -- I was under the impression that preempt_count() covered
> this as well, due to the following in include/linux/hardirq.h:
> 
> 	#define PREEMPT_MASK    (__IRQ_MASK(PREEMPT_BITS) << PREEMPT_SHIFT)
> 	#define SOFTIRQ_MASK    (__IRQ_MASK(SOFTIRQ_BITS) << SOFTIRQ_SHIFT)
> 	#define HARDIRQ_MASK    (__IRQ_MASK(HARDIRQ_BITS) << HARDIRQ_SHIFT)
> 	#define NMI_MASK        (__IRQ_MASK(NMI_BITS)     << NMI_SHIFT)
> 
> But irqs_disabled() does look to sample the actual interrupt hardware.
> 
> So, if there are cases where RCU is used where the interrupt hardware
> is disabled, but preempt_count() has not been updated, this patch is
> the right thing to do.
> 
> 							Thanx, Paul
> 

local_irq_disable() does not touch preempt_count, it touchs
a register of current CPU.

The following quick test module is a case where RCU_SCHED is used where
the interrupt hardware is disabled, but preempt_count() has
not been updated, and it raises rcu-lockdep complaint.

#include <linux/module.h>
#include <linux/hardirq.h>
#include <linux/rcupdate.h>

void *test = &test;

int test_init(void)
{
	local_irq_disable();
	printk(KERN_INFO "%p\n", rcu_dereference_sched(test));
	local_irq_enable();

	return 0;
}

void test_exit(void) {}

module_init(test_init);
module_exit(test_exit);
MODULE_LICENSE("GPL");