[Intel-gfx] [PATCH v2 04/14] vfio: Provide better generic support for open/release vfio_device_ops

[Jason Gunthorpe]

On Fri, Jul 23, 2021 at 09:39:14AM +0200, Christoph Hellwig wrote:

> This looks unessecarily complicated.  We can just try to load first
> and then store it under the same lock, e.g.:

Yes indeed, I went with this:

int vfio_assign_device_set(struct vfio_device *device, void *set_id)
{
	unsigned long idx = (unsigned long)set_id;
	struct vfio_device_set *new_dev_set;
	struct vfio_device_set *dev_set;

	if (WARN_ON(!set_id))
		return -EINVAL;

	/*
	 * Atomically acquire a singleton object in the xarray for this set_id
	 */
	xa_lock(&vfio_device_set_xa);
	dev_set = xa_load(&vfio_device_set_xa, idx);
	if (dev_set)
		goto found_get_ref;
	xa_unlock(&vfio_device_set_xa);

	new_dev_set = kzalloc(sizeof(*new_dev_set), GFP_KERNEL);
	if (!new_dev_set)
		return -ENOMEM;
	mutex_init(&new_dev_set->lock);
	new_dev_set->set_id = set_id;

	xa_lock(&vfio_device_set_xa);
	dev_set = __xa_cmpxchg(&vfio_device_set_xa, idx, NULL, new_dev_set,
			       GFP_KERNEL);
	if (!dev_set) {
		dev_set = new_dev_set;
		goto found_get_ref;
	}

	kfree(new_dev_set);
	if (xa_is_err(dev_set)) {
		xa_unlock(&vfio_device_set_xa);
		return xa_err(dev_set);
	}

found_get_ref:
	dev_set->device_count++;
	xa_unlock(&vfio_device_set_xa);
	device->dev_set = dev_set;
	return 0;
}

I'm also half inclined to delete the xa_load() since I think the
common case here is to need the allocate...

> 	xa_lock(&vfio_device_set_xa);
> 	set = xa_load(&vfio_device_set_xa, idx);
> 	if (set) {
> 		kfree(new);
> 		goto found;
> 	}
> 	err = xa_err(__xa_store(&vfio_device_set_xa, idx, new, GFP_KERNEL));

AIUI this is subtly racy:


  CPU1                               CPU2
xa_lock()
 xa_load() == NULL
 xa_store()
   __xas_nomem()
     xa_unlock()
                                 xa_lock()
                                  xa_load() == NULL
                                  xa_store()
                                   __xas_nomem()
                                    xa_unlock()
                                     kmem_cache_alloc()
      kmem_cache_alloc()
     xa_lock()
   [idx] = new1
xa_unlock()
                                    xa_lock()
                                   [idx] = new2    // Woops, lost new1!
                                 xa_unlock()

The hidden xa unlock is really tricky.

The __xa_cmpxchg is safe against this.

Jason