Skip to contentSkip to conversation

CPU ordering with respect to krefs

PATCH Oliver Neukum 3 participants 4 messages Apr 2, 2007 – Apr 17, 2007 Library code

lore
Share & tools
lore
#0 lore raw

Hi,

some atomic operations are only atomic, not ordered. Thus a CPU is allowed to reorder memory references to an object to before the reference is obtained. This fixes it.

Regards
Oliver

------

lib/kref.c +2 -0
@@ -21,6 +21,7 @@
 void kref_init(struct kref *kref)
 {
 	atomic_set(&kref->refcount,1);
+	smp_mb();
 }
 
 /**
@@ -31,6 +32,7 @@
 {
 	WARN_ON(!atomic_read(&kref->refcount));
 	atomic_inc(&kref->refcount);
+	smp_mb__after_atomic_inc();
 }
 
 /**
#1 lore raw
Quoted 14 lines · On Mon, 2 Apr 2007 14:47:59 +0200 Oliver Neukum <[email protected]> wrote:

Hi,

some atomic operations are only atomic, not ordered. Thus a CPU is allowed to reorder memory references to an object to before the reference is obtained. This fixes it.

Regards
Oliver

------

lib/kref.c line 25

Quoted 5 lines
lib/kref.c +1 -0
@@ -21,6 +21,7 @@
 void kref_init(struct kref *kref)
 {
 	atomic_set(&kref->refcount,1);
+	smp_mb();
 }

I dont understand why smp_mb() is needed here, and not in spinlock_init() for example.

If you have ordering issues, then the caller of kref_init() should take care of it, not kref_init() itself.

Random example taken in drivers/usb/gadget/file_storage.c :

static int __init fsg_alloc(void)
{
...
kref_init(&fsg->ref);
init_completion(&fsg->thread_notifier);

the_fsg = fsg;
}

In this example, "the_fsg = fsg" memory write might be visible before the memory writes done in init_completion(). Doing a smp_mb() in kref_init() wont help.

AFAIK kref implementation doesnt need this extra smp_mb().

#2 lore raw
Quoted 14 lines · Eric Dumazet
Quoted 11 lines · On Mon, 2 Apr 2007 14:47:59 +0200 Oliver Neukum <[email protected]> wrote:

Hi,

some atomic operations are only atomic, not ordered. Thus a CPU is allowed to reorder memory references to an object to before the reference is obtained. This fixes it.

Regards
Oliver

------

Quoted 11 lines

lib/kref.c line 25

Quoted 8 lines
lib/kref.c +1 -0
@@ -21,6 +21,7 @@
 void kref_init(struct kref *kref)
 {
 	atomic_set(&kref->refcount,1);
+	smp_mb();
 }

I dont understand why smp_mb() is needed here, and not in spinlock_init() for example.

I think, after reading the Documentation/memory-barriers.txt and Documentation/atomic_ops.txt documentation, that spin_lock_init() also needs this kind of memory barrier.

Quoted 1 line

From what I can tell (Oliver, please correct me if I'm wrong, you know

this much better than I do), the issue is that atomic init has no memory barrier, and you need to handle that.

thanks,

greg k-h

#3 lore raw

Am Donnerstag, 12. April 2007 08:27 schrieb Greg KH:

Quoted 27 lines · Eric Dumazet
Quoted 14 lines
Quoted 11 lines · On Mon, 2 Apr 2007 14:47:59 +0200 Oliver Neukum <[email protected]> wrote:

Hi,

some atomic operations are only atomic, not ordered. Thus a CPU is allowed to reorder memory references to an object to before the reference is obtained. This fixes it.

Regards
Oliver

------

Quoted 11 lines

lib/kref.c line 25

Quoted 8 lines
lib/kref.c +1 -0
@@ -21,6 +21,7 @@
 void kref_init(struct kref *kref)
 {
 	atomic_set(&kref->refcount,1);
+	smp_mb();
 }

I dont understand why smp_mb() is needed here, and not in spinlock_init() for example.

Quoted 3 lines

I think, after reading the Documentation/memory-barriers.txt and Documentation/atomic_ops.txt documentation, that spin_lock_init() also needs this kind of memory barrier.

spin_lock_init() is not an atomic operation.
In principle, the issue exists. However, the whole issue is a bit of a grey area. You might take the viewpoint that upping the refcount needs to be under lock, which needs to take care of ordering issues in case of krefs. A new spinlock has the same issue. You need to be careful making them accessible to other CPUs.

If you take code like:

static int producer()

{
	...
	data = kmalloc(...);
	spin_lock_init(&data->lock);
	data->value = some_value;
	data->next = global_pointer;

	global_pointer = data;
	...
}

You have an ordering bug anyway, which you can't fix in spin_lock_init().

Regards
Oliver