Re: [PATCH] sys_getppid oopses on debug kernel (v2)

Re: [PATCH] sys_getppid oopses on debug kernel (v2)

[Oleg Nesterov]

Andrew Morton wrote:
>
> Although I'm not sure it's needed for this problem. A getppid() which does
>
> asmlinkage long sys_getppid(void)
> {
> 	int pid;
>
> 	read_lock(&tasklist_lock);
> 	pid = current->group_leader->real_parent->tgid;
> 	read_unlock(&tasklist_lock);
>
> 	return pid;
> }
>
> seems like a fine implementation to me ;)

Why do we need to use ->group_leader? All threads should have the same
->real_parent.

Why do we need tasklist_lock? I think rcu_read_lock() is enough.

In other words, do you see any problems with this code

	smlinkage long sys_getppid(void)
	{
		int pid;

		rcu_read_lock();
		pid = rcu_dereference(current->real_parent)->tgid;
		rcu_read_unlock();

		return pid;
	}

? Yes, we may read a stale value for ->real_parent, but the memory
can't be freed while we are under rcu_read_lock(). And in this case
the returned value is ok because the task could be reparented just
after return anyway.

Oleg.

Re: [PATCH] sys_getppid oopses on debug kernel (v2)

[Kirill Korotaev]

>>Although I'm not sure it's needed for this problem. A getppid() which does
>>
>>asmlinkage long sys_getppid(void)
>>{
>>	int pid;
>>
>>	read_lock(&tasklist_lock);
>>	pid = current->group_leader->real_parent->tgid;
>>	read_unlock(&tasklist_lock);
>>
>>	return pid;
>>}
>>
>>seems like a fine implementation to me ;)
> 
> 
> Why do we need to use ->group_leader? All threads should have the same
> ->real_parent.
I'm not sure this is true for old LinuxThreads...

> Why do we need tasklist_lock? I think rcu_read_lock() is enough.
> 
> In other words, do you see any problems with this code
> 
> 	smlinkage long sys_getppid(void)
> 	{
> 		int pid;
> 
> 		rcu_read_lock();
> 		pid = rcu_dereference(current->real_parent)->tgid;
> 		rcu_read_unlock();
> 
> 		return pid;
> 	}
> 
> ? Yes, we may read a stale value for ->real_parent, but the memory
> can't be freed while we are under rcu_read_lock(). And in this case
> the returned value is ok because the task could be reparented just
> after return anyway.
Your patch doesn't cure the problem.
rcu_read_lock just disables preemtion and rcu_dereference
introduces memory barrier. _None_ of this _prevents_
another CPU from freeing old real_parent in parallel with your dereference.

You can minimize the probability very much by making local_irq_disable()/enable()
around the code in question, but still it won't be a real fix (at least due to NMIs).

Kirill

Re: [PATCH] sys_getppid oopses on debug kernel (v2)

[Oleg Nesterov]

On 08/09, Kirill Korotaev wrote:
>
> >Why do we need to use ->group_leader? All threads should have the same
> >->real_parent.
> I'm not sure this is true for old LinuxThreads...

Yes, from the user-space pov it may be not true, but ->group_leader->real_parent
should be equal ->real_parent anyway.

> >Why do we need tasklist_lock? I think rcu_read_lock() is enough.
> >
> >In other words, do you see any problems with this code
> >
> >	smlinkage long sys_getppid(void)
> >	{
> >		int pid;
> >
> >		rcu_read_lock();
> >		pid = rcu_dereference(current->real_parent)->tgid;
> >		rcu_read_unlock();
> >
> >		return pid;
> >	}
> >
> >? Yes, we may read a stale value for ->real_parent, but the memory
> >can't be freed while we are under rcu_read_lock(). And in this case
> >the returned value is ok because the task could be reparented just
> >after return anyway.
> Your patch doesn't cure the problem.
> rcu_read_lock just disables preemtion and rcu_dereference
> introduces memory barrier. _None_ of this _prevents_
> another CPU from freeing old real_parent in parallel with your dereference.

How so? Note that release_task() doesn't call put_task_struct(), it does
call_rcu(&p->rcu, delayed_put_task_struct) instead. When delayed_put_task_struct()
is called, all CPUs must see the new value of ->real_parent (otherwise
RCU is just broken). If CPU sees the old value of ->real_parent, rcu_read_lock()
protects us from delayed_put_task_struct() on another CPU.

Ok, I think this is the same "classic" pattern as:

	old = global_ptr;
	global_ptr = new;
	call_rcu(..free_old...);
vs
	rcu_read_lock();
	use(global_ptr);
	rcu_read_unlock();

Do you agree?

Oleg.

Re: [PATCH] sys_getppid oopses on debug kernel (v2)

[Kirill Korotaev]

>>Your patch doesn't cure the problem.
>>rcu_read_lock just disables preemtion and rcu_dereference
>>introduces memory barrier. _None_ of this _prevents_
>>another CPU from freeing old real_parent in parallel with your dereference.
> 
> 
> How so? Note that release_task() doesn't call put_task_struct(), it does
> call_rcu(&p->rcu, delayed_put_task_struct) instead. When delayed_put_task_struct()
> is called, all CPUs must see the new value of ->real_parent (otherwise
> RCU is just broken). If CPU sees the old value of ->real_parent, rcu_read_lock()
> protects us from delayed_put_task_struct() on another CPU.
> 
> Ok, I think this is the same "classic" pattern as:
> 
> 	old = global_ptr;
> 	global_ptr = new;
> 	call_rcu(..free_old...);
> vs
> 	rcu_read_lock();
> 	use(global_ptr);
> 	rcu_read_unlock();
> 
> Do you agree?
Agree :)
Sorry, I didn't notice that task structs are freed now with RCU :)))
In this case your patch really helps the problem. Care to prepare it?

Thanks,
Kirill

[PATCH] sys_getppid-oopses-on-debug-kernel-v2-simplify

[Oleg Nesterov]

On top of Kirill's sys_getppid-oopses-on-debug-kernel-v2.patch

- We don't need ->group_leader->real_parent, all threads should
  have the same ->real_parent.

- We don't need tasklist_lock, task_struct is freed by RCU, so
  rcu_read_lock() should be enough.

(Compile tested)

Signed-off-by: Oleg Nesterov <oleg@tv-sign.ru>

--- 2.6.18-rc3/kernel/timer.c~	2006-08-09 22:08:51.000000000 +0400
+++ 2.6.18-rc3/kernel/timer.c	2006-08-09 22:15:35.000000000 +0400
@@ -1324,28 +1324,18 @@ asmlinkage long sys_getpid(void)
 }
 
 /*
- * Accessing ->group_leader->real_parent is not SMP-safe, it could
- * change from under us. However, rather than getting any lock
- * we can use an optimistic algorithm: get the parent
- * pid, and go back and check that the parent is still
- * the same. If it has changed (which is extremely unlikely
- * indeed), we just try again..
- *
- * NOTE! This depends on the fact that even if we _do_
- * get an old value of "parent", we can happily dereference
- * the pointer (it was and remains a dereferencable kernel pointer
- * no matter what): we just can't necessarily trust the result
- * until we know that the parent pointer is valid.
- *
- * NOTE2: ->group_leader never changes from under us.
+ * Accessing ->real_parent is not SMP-safe, it could
+ * change from under us. However, we can use a stale
+ * value of ->real_parent under rcu_read_lock(), see
+ * release_task()->call_rcu(delayed_put_task_struct).
  */
 asmlinkage long sys_getppid(void)
 {
 	int pid;
 
-	read_lock(&tasklist_lock);
-	pid = current->group_leader->real_parent->tgid;
-	read_unlock(&tasklist_lock);
+	rcu_read_lock();
+	pid = rcu_dereference(current->real_parent)->tgid;
+	rcu_read_unlock();
 
 	return pid;
 }