Re: [PATCH v9 09/22] s390: vfio-ap: register matrix device with VFIO mdev framework

[Cornelia Huck <cohuck@redhat.com>]

On Thu, 16 Aug 2018 12:24:16 -0400
Tony Krowiak <akrowiak@linux.ibm.com> wrote:

> On 08/14/2018 07:19 AM, Cornelia Huck wrote:
> > On Mon, 13 Aug 2018 17:48:06 -0400
> > Tony Krowiak <akrowiak@linux.vnet.ibm.com> wrote:

> >> +static int vfio_ap_mdev_create(struct kobject *kobj, struct mdev_device *mdev)
> >> +{
> >> +	struct ap_matrix_mdev *matrix_mdev;
> >> +
> >> +	matrix_mdev = kzalloc(sizeof(*matrix_mdev), GFP_KERNEL);
> >> +	if (!matrix_mdev)
> >> +		return -ENOMEM;
> >> +
> >> +	matrix_mdev->name = dev_name(mdev_dev(mdev));
> >> +	vfio_ap_matrix_init(&matrix_dev.info, &matrix_mdev->matrix);
> >> +	mdev_set_drvdata(mdev, matrix_mdev);
> >> +
> >> +	if (atomic_dec_if_positive(&matrix_dev.available_instances) < 0) {
> >> +		kfree(matrix_mdev);
> >> +		return -EPERM;
> >> +	}  
> > Maybe move this check to the top of the function?  
> 
> Please ignore my previous response to your comment. I can't move the call to
> atomic_dec_if_positive() to the top of the function because it 
> decrements the
> available_instances and if the kzalloc() of matrix_mdev fails, then the 
> value
> would have to then be incremented to remain valid. What I can do is this:
> 
> 1. Check the value of available_instances using atomic_read() at the top of
>     the function and if it is zero, return an error.
> 
> 2. Replace the call to atomic_dec_if_positive() with a call to atomic_dec()
>     to decrement the available_instances.
> 
> I agree that it makes sense to return before attempting to allocate the
> matrix_mdev if available_instances is zero.

Wouldn't that be racy, though?

I don't think re-incrementing the counter is too bad, and it's
certainly better than going through allocation/freeing of structures.

> 
> >  
> >> +
> >> +	mutex_lock(&matrix_dev.lock);
> >> +	list_add(&matrix_mdev->list, &matrix_dev.mdev_list);
> >> +	mutex_unlock(&matrix_dev.lock);
> >> +
> >> +	return 0;
> >> +}  
> 
>