Re: lockdep rcu-preempt and synchronize_srcu() awareness

["Paul E. McKenney" <paulmck@linux.vnet.ibm.com>]

On Mon, Feb 08, 2010 at 02:18:58PM -0500, Mathieu Desnoyers wrote:
> Hi,
> 
> I just though about the following deadlock scenario involving rcu preempt and
> mutexes. I see that lockdep does not warn about it, and it actually triggers a
> deadlock on my box. It might be worth addressing for TREE_PREEMPT_RCU configs.
> 
> CPU A:
>     mutex_lock(&test_mutex);
>     synchronize_rcu();
>     mutex_unlock(&test_mutex);
> 
> CPU B:
>     rcu_read_lock();
>     mutex_lock(&test_mutex);
>     mutex_unlock(&test_mutex);
>     rcu_read_unlock();
> 
> But given that it's not legit to take a mutex from within a rcu read lock in
> non-preemptible configs, I guess it's not much of a real-life problem, but I
> think SRCU is also affected, because there is no lockdep annotation around
> synchronize_srcu().

Indeed, doing this with SRCU would result in deadlock, and it is quite
legal to acquire mutexes from within SRCU read-side critical sections.
And similar deadlocks can be constructed using pthread_mutex_lock() and
user-space RCU implementations.

The basic rule is "don't wait for a grace period to complete while in
the corresponding flavor of RCU read-side critical section".  Your point,
that it is possible to wait indirectly, is well taken.

> So I think it would be good to mark rcu_read_lock/unlock as not permitting
> "might_sleep()" in non preemptable RCU configs, and having a look at lockdep
> SRCU support might be worthwhile.

Given the in-progress lockdep enhancements to RCU, the information is at
least present.  I can easily check for the direct case, but must defer
to Peter Z on the indirect case.

							Thanx, Paul

> The following test module triggers the problem:
> 
> 
> /* test-rcu-lockdep.c
>  *
>  * Test RCU-awareness of lockdep. Don't look at the interface, it's aweful.
>  * run, in parallel:
>  *
>  * cat /proc/testa 
>  * cat /proc/testb
>  */
> 
> #include <linux/module.h>
> #include <linux/mutex.h>
> #include <linux/proc_fs.h>
> #include <linux/sched.h>
> #include <linux/delay.h>
> 
> struct proc_dir_entry *pentrya = NULL;
> struct proc_dir_entry *pentryb = NULL;
> 
> static DEFINE_MUTEX(test_mutex);
> 
> static int my_opena(struct inode *inode, struct file *file)
> {
> 	mutex_lock(&test_mutex);
> 	synchronize_rcu();
> 	mutex_unlock(&test_mutex);
> 
> 	return -EPERM;
> }
> 
> 
> static struct file_operations my_operationsa = {
> 	.open = my_opena,
> };
> 
> static int my_openb(struct inode *inode, struct file *file)
> {
> 	rcu_read_lock();
> 	mutex_lock(&test_mutex);
> 	ssleep(1);
> 	mutex_unlock(&test_mutex);
> 	rcu_read_unlock();
> 
> 
> 	return -EPERM;
> }
> 
> 
> static struct file_operations my_operationsb = {
> 	.open = my_openb,
> };
> 
> int init_module(void)
> {
> 	pentrya = create_proc_entry("testa", 0444, NULL);
> 	if (pentrya)
> 		pentrya->proc_fops = &my_operationsa;
> 
> 	pentryb = create_proc_entry("testb", 0444, NULL);
> 	if (pentryb)
> 		pentryb->proc_fops = &my_operationsb;
> 
> 	return 0;
> }
> 
> void cleanup_module(void)
> {
> 	remove_proc_entry("testa", NULL);
> 	remove_proc_entry("testb", NULL);
> }
> 
> MODULE_LICENSE("GPL");
> MODULE_AUTHOR("Mathieu Desnoyers");
> MODULE_DESCRIPTION("lockdep rcu test");
> 
> 
> 
> Thanks,
> 
> Mathieu