Re: [PATCH v3 1/1] s390/vfio-ap: fix circular lockdep when setting/clearing crypto masks

[Tony Krowiak <akrowiak@linux.ibm.com>]

On 3/3/21 2:47 PM, Halil Pasic wrote:
> On Wed, 3 Mar 2021 12:10:11 -0500
> Tony Krowiak <akrowiak@linux.ibm.com> wrote:
>
>> On 3/3/21 10:23 AM, Halil Pasic wrote:
>>> On Tue,  2 Mar 2021 15:43:22 -0500
>>> Tony Krowiak <akrowiak@linux.ibm.com> wrote:
>>>   
>>>> This patch fixes a lockdep splat introduced by commit f21916ec4826
>>>> ("s390/vfio-ap: clean up vfio_ap resources when KVM pointer invalidated").
>>>> The lockdep splat only occurs when starting a Secure Execution guest.
>>>> Crypto virtualization (vfio_ap) is not yet supported for SE guests;
>>>> however, in order to avoid this problem when support becomes available,
>>>> this fix is being provided.
>>> [..]
>>>   
>>>> @@ -1038,14 +1116,28 @@ static int vfio_ap_mdev_set_kvm(struct ap_matrix_mdev *matrix_mdev,
>>>>    {
>>>>    	struct ap_matrix_mdev *m;
>>>>
>>>> -	list_for_each_entry(m, &matrix_dev->mdev_list, node) {
>>>> -		if ((m != matrix_mdev) && (m->kvm == kvm))
>>>> -			return -EPERM;
>>>> -	}
>>>> +	if (kvm->arch.crypto.crycbd) {
>>>> +		matrix_mdev->kvm_busy = true;
>>>>
>>>> -	matrix_mdev->kvm = kvm;
>>>> -	kvm_get_kvm(kvm);
>>>> -	kvm->arch.crypto.pqap_hook = &matrix_mdev->pqap_hook;
>>>> +		list_for_each_entry(m, &matrix_dev->mdev_list, node) {
>>>> +			if ((m != matrix_mdev) && (m->kvm == kvm)) {
>>>> +				wake_up_all(&matrix_mdev->wait_for_kvm);
>>> This ain't no good. kvm_busy will remain true if we take this exit. The
>>> wake_up_all() is not needed, because we hold the lock, so nobody can
>>> observe it if we don't forget kvm_busy set.
>>>
>>> I suggest moving matrix_mdev->kvm_busy = true; after this loop, maybe right
>>> before the unlock, and removing the wake_up_all().
>>>   
>>>> +				return -EPERM;
>>>> +			}
>>>> +		}
>>>> +
>>>> +		kvm_get_kvm(kvm);
>>>> +		mutex_unlock(&matrix_dev->lock);
>>>> +		kvm_arch_crypto_set_masks(kvm,
>>>> +					  matrix_mdev->matrix.apm,
>>>> +					  matrix_mdev->matrix.aqm,
>>>> +					  matrix_mdev->matrix.adm);
>>>> +		mutex_lock(&matrix_dev->lock);
>>>> +		kvm->arch.crypto.pqap_hook = &matrix_mdev->pqap_hook;
>>>> +		matrix_mdev->kvm = kvm;
>>>> +		matrix_mdev->kvm_busy = false;
>>>> +		wake_up_all(&matrix_mdev->wait_for_kvm);
>>>> +	}
>>>>
>>>>    	return 0;
>>>>    }
>>> [..]
>>>   
>>>> @@ -1300,7 +1406,21 @@ static ssize_t vfio_ap_mdev_ioctl(struct mdev_device *mdev,
>>>>    		ret = vfio_ap_mdev_get_device_info(arg);
>>>>    		break;
>>>>    	case VFIO_DEVICE_RESET:
>>>> -		ret = vfio_ap_mdev_reset_queues(mdev);
>>>> +		matrix_mdev = mdev_get_drvdata(mdev);
>>>> +
>>>> +		/*
>>>> +		 * If the KVM pointer is in the process of being set, wait until
>>>> +		 * the process has completed.
>>>> +		 */
>>>> +		wait_event_cmd(matrix_mdev->wait_for_kvm,
>>>> +			       matrix_mdev->kvm_busy == false,
>>>> +			       mutex_unlock(&matrix_dev->lock),
>>>> +			       mutex_lock(&matrix_dev->lock));
>>>> +
>>>> +		if (matrix_mdev->kvm)
>>>> +			ret = vfio_ap_mdev_reset_queues(mdev);
>>>> +		else
>>>> +			ret = -ENODEV;
>>> I don't think rejecting the reset is a good idea. I have you a more detailed
>>> explanation of the list, where we initially discussed this question.
>>>
>>> How do you exect userspace to react to this -ENODEV?
>> After reading your more detailed explanation, I have come to the
>> conclusion that the test for matrix_mdev->kvm should not be
>> performed here and the the vfio_ap_mdev_reset_queues() function
>> should be called regardless. Each queue assigned to the mdev
>> that is also bound to the vfio_ap driver will get reset and its
>> IRQ resources cleaned up if they haven't already been and the
>> other required conditions are met (i.e., see
>> vfio_ap_mdev_free_irq_resources()).
> My point is if !->kvm the other required conditions are not met. But
> yes we can go back to unconditional vfio_ap_mdev_reset_queues(mdev),
> and think about the necessity of performing a
> vfio_ap_mdev_reset_queues() if !->kvm later as I proposed in the other
> mail.

The other conditions may or may not have been met depending
upon whether ->kvm is NULL because the VFIO_DEVICE_RESET
ioctl was invoked while the matrix_dev->lock was released
in the vfio_ap_mdev_unset_kvm() function. If that was the case,
then there is no need to clean up the IRQ resources because it
will already have been done.

On the other hand, if we don't have ->kvm because something broke,
then we may be out of luck anyway. There will certainly be no
way to unregister the GISC; however, it may still be possible
to unpin the pages if we still have q->saved_pfn.

The point is, if the queue is bound to vfio_ap, it can be reset. If we can't
clean up the IRQ resources because something is broken, then there
is nothing we can do about that.


>
> Regards,
> Halil