[PATCH -mm 2/2] is_current_single_threaded: don't use ->mmap_sem

[PATCH -mm 2/2] is_current_single_threaded: don't use ->mmap_sem

[Oleg Nesterov]

is_current_single_threaded() can safely miss a freshly forked CLONE_VM
task, but in this case it must not miss its parent. That is why we take
mm->mmap_sem for writing to make sure a thread/task with the same ->mm
can't pass exit_mm() and disappear.

However we can avoid ->mmap_sem and rely on rcu/barriers:

	- if we do not see the exiting parent on thread/process list
	  we see the result of list_del_rcu(), in this case we must
	  also see the result of list_add_rcu() which does wmb().

	- if we do see the parent but its ->mm == NULL, we need rmb()
	  to make sure we can't miss the child.

Signed-off-by: Oleg Nesterov <oleg@redhat.com>
---

 lib/is_single_threaded.c |   10 ++++++----
 1 file changed, 6 insertions(+), 4 deletions(-)

--- WAIT/lib/is_single_threaded.c~ISS_2_RCU	2009-07-09 22:16:29.000000000 +0200
+++ WAIT/lib/is_single_threaded.c	2009-07-09 22:54:41.000000000 +0200
@@ -22,8 +22,6 @@ bool is_current_single_threaded(void)
 	struct task_struct *p, *t;
 	bool ret;
 
-	might_sleep();
-
 	if (atomic_read(&task->signal->live) != 1)
 		return false;
 
@@ -31,7 +29,6 @@ bool is_current_single_threaded(void)
 		return true;
 
 	ret = false;
-	down_write(&mm->mmap_sem);
 	rcu_read_lock();
 	for_each_process(p) {
 		if (unlikely(p->flags & PF_KTHREAD))
@@ -45,12 +42,17 @@ bool is_current_single_threaded(void)
 				goto found;
 			if (likely(t->mm))
 				break;
+			/*
+			 * t->mm == NULL. Make sure next_thread/next_task
+			 * will see other CLONE_VM tasks which might be
+			 * forked before exiting.
+			 */
+			smp_rmb();
 		} while_each_thread(p, t);
 	}
 	ret = true;
 found:
 	rcu_read_unlock();
-	up_write(&mm->mmap_sem);
 
 	return ret;
 }

Re: [PATCH -mm 2/2] is_current_single_threaded: don't use ->mmap_sem

[Oleg Nesterov]

On 07/09, Oleg Nesterov wrote:
>
> is_current_single_threaded() can safely miss a freshly forked CLONE_VM
> task, but in this case it must not miss its parent. That is why we take
> mm->mmap_sem for writing to make sure a thread/task with the same ->mm
> can't pass exit_mm() and disappear.
>
> However we can avoid ->mmap_sem and rely on rcu/barriers:
>
> 	- if we do not see the exiting parent on thread/process list
> 	  we see the result of list_del_rcu(), in this case we must
> 	  also see the result of list_add_rcu() which does wmb().
>
> 	- if we do see the parent but its ->mm == NULL, we need rmb()
> 	  to make sure we can't miss the child.

To simplify the review, below is the code after the patch.

Oleg.

/*
 * Returns true if the task does not share ->mm with another thread/process.
 */
bool is_current_single_threaded(void)
{
	struct task_struct *task = current;
	struct mm_struct *mm = task->mm;
	struct task_struct *p, *t;
	bool ret;

	if (atomic_read(&task->signal->live) != 1)
		return false;

	if (atomic_read(&mm->mm_users) == 1)
		return true;

	ret = false;
	rcu_read_lock();
	for_each_process(p) {
		if (unlikely(p->flags & PF_KTHREAD))
			continue;
		if (unlikely(p == task->group_leader))
			continue;

		t = p;
		do {
			if (unlikely(t->mm == mm))
				goto found;
			if (likely(t->mm))
				break;
			/*
			 * t->mm == NULL. Make sure next_thread/next_task
			 * will see other CLONE_VM tasks which might be
			 * forked before exiting.
			 */
			smp_rmb();
		} while_each_thread(p, t);
	}
	ret = true;
found:
	rcu_read_unlock();

	return ret;
}