[RFC PATCH v1] fixup! s390/vfio-ap: Notify userspace that guest's AP config changed when mdev removed

[RFC PATCH v1] fixup! s390/vfio-ap: Notify userspace that guest's AP config changed when mdev removed

[Rorie Reyes]

This patch is based on the s390/features branch

The guest's AP configuration is cleared when the mdev is removed, so
userspace must be notified that the AP configuration has changed. To this
end, this patch:

* Removes call to 'signal_guest_ap_cfg_changed()' function from the
  'vfio_ap_mdev_unset_kvm()' function because it has no affect given it is
  called after the mdev fd is closed.

* Adds call to 'signal_guest_ap_cfg_changed()' function to the
  'vfio_ap_mdev_request()' function to notify userspace that the guest's
  AP configuration has changed before signaling the request to remove the
  mdev.

Minor change - Fixed an indentation issue in function
'signal_guest_ap_cfg_changed()'

Fixes: 2ba4410dd477 ("s390/vfio-ap: Signal eventfd when guest AP configuration is changed")
Signed-off-by: Rorie Reyes <rreyes@linux.ibm.com>
---
 drivers/s390/crypto/vfio_ap_ops.c | 19 ++++++++++++++++---
 1 file changed, 16 insertions(+), 3 deletions(-)

diff --git a/drivers/s390/crypto/vfio_ap_ops.c b/drivers/s390/crypto/vfio_ap_ops.c
index 571f5dcb49c5..c1afac5ac555 100644
--- a/drivers/s390/crypto/vfio_ap_ops.c
+++ b/drivers/s390/crypto/vfio_ap_ops.c
@@ -652,8 +652,8 @@ static void vfio_ap_matrix_init(struct ap_config_info *info,
 
 static void signal_guest_ap_cfg_changed(struct ap_matrix_mdev *matrix_mdev)
 {
-		if (matrix_mdev->cfg_chg_trigger)
-			eventfd_signal(matrix_mdev->cfg_chg_trigger);
+	if (matrix_mdev->cfg_chg_trigger)
+		eventfd_signal(matrix_mdev->cfg_chg_trigger);
 }
 
 static void vfio_ap_mdev_update_guest_apcb(struct ap_matrix_mdev *matrix_mdev)
@@ -1870,7 +1870,6 @@ static void vfio_ap_mdev_unset_kvm(struct ap_matrix_mdev *matrix_mdev)
 		get_update_locks_for_kvm(kvm);
 
 		kvm_arch_crypto_clear_masks(kvm);
-		signal_guest_ap_cfg_changed(matrix_mdev);
 		vfio_ap_mdev_reset_queues(matrix_mdev);
 		kvm_put_kvm(kvm);
 		matrix_mdev->kvm = NULL;
@@ -2057,6 +2056,14 @@ static void vfio_ap_mdev_request(struct vfio_device *vdev, unsigned int count)
 
 	matrix_mdev = container_of(vdev, struct ap_matrix_mdev, vdev);
 
+	if (matrix_mdev->kvm) {
+		get_update_locks_for_kvm(matrix_mdev->kvm);
+		kvm_arch_crypto_clear_masks(matrix_mdev->kvm);
+		signal_guest_ap_cfg_changed(matrix_mdev);
+	} else {
+		mutex_lock(&matrix_dev->mdevs_lock);
+	}
+
 	if (matrix_mdev->req_trigger) {
 		if (!(count % 10))
 			dev_notice_ratelimited(dev,
@@ -2068,6 +2075,12 @@ static void vfio_ap_mdev_request(struct vfio_device *vdev, unsigned int count)
 		dev_notice(dev,
 			   "No device request registered, blocked until released by user\n");
 	}
+
+	if (matrix_mdev->kvm)
+		release_update_locks_for_kvm(matrix_mdev->kvm);
+	else
+		mutex_unlock(&matrix_dev->mdevs_lock);
+
 }
 
 static int vfio_ap_mdev_get_device_info(unsigned long arg)
-- 
2.48.1

Re: [RFC PATCH v1] fixup! s390/vfio-ap: Notify userspace that guest's AP config changed when mdev removed

[Anthony Krowiak]

On 3/3/25 2:11 PM, Rorie Reyes wrote:
> This patch is based on the s390/features branch
>
> The guest's AP configuration is cleared when the mdev is removed, so
> userspace must be notified that the AP configuration has changed. To this
> end, this patch:
>
> * Removes call to 'signal_guest_ap_cfg_changed()' function from the
>    'vfio_ap_mdev_unset_kvm()' function because it has no affect given it is
>    called after the mdev fd is closed.
>
> * Adds call to 'signal_guest_ap_cfg_changed()' function to the
>    'vfio_ap_mdev_request()' function to notify userspace that the guest's
>    AP configuration has changed before signaling the request to remove the
>    mdev.
>
> Minor change - Fixed an indentation issue in function
> 'signal_guest_ap_cfg_changed()'
>
> Fixes: 2ba4410dd477 ("s390/vfio-ap: Signal eventfd when guest AP configuration is changed")
> Signed-off-by: Rorie Reyes <rreyes@linux.ibm.com>
> ---
>   drivers/s390/crypto/vfio_ap_ops.c | 19 ++++++++++++++++---
>   1 file changed, 16 insertions(+), 3 deletions(-)
>
> diff --git a/drivers/s390/crypto/vfio_ap_ops.c b/drivers/s390/crypto/vfio_ap_ops.c
> index 571f5dcb49c5..c1afac5ac555 100644
> --- a/drivers/s390/crypto/vfio_ap_ops.c
> +++ b/drivers/s390/crypto/vfio_ap_ops.c
> @@ -652,8 +652,8 @@ static void vfio_ap_matrix_init(struct ap_config_info *info,
>   
>   static void signal_guest_ap_cfg_changed(struct ap_matrix_mdev *matrix_mdev)
>   {
> -		if (matrix_mdev->cfg_chg_trigger)
> -			eventfd_signal(matrix_mdev->cfg_chg_trigger);
> +	if (matrix_mdev->cfg_chg_trigger)
> +		eventfd_signal(matrix_mdev->cfg_chg_trigger);
>   }
>   
>   static void vfio_ap_mdev_update_guest_apcb(struct ap_matrix_mdev *matrix_mdev)
> @@ -1870,7 +1870,6 @@ static void vfio_ap_mdev_unset_kvm(struct ap_matrix_mdev *matrix_mdev)
>   		get_update_locks_for_kvm(kvm);
>   
>   		kvm_arch_crypto_clear_masks(kvm);
> -		signal_guest_ap_cfg_changed(matrix_mdev);
>   		vfio_ap_mdev_reset_queues(matrix_mdev);
>   		kvm_put_kvm(kvm);
>   		matrix_mdev->kvm = NULL;
> @@ -2057,6 +2056,14 @@ static void vfio_ap_mdev_request(struct vfio_device *vdev, unsigned int count)
>   
>   	matrix_mdev = container_of(vdev, struct ap_matrix_mdev, vdev);
>   
> +	if (matrix_mdev->kvm) {
> +		get_update_locks_for_kvm(matrix_mdev->kvm);

I know we talked about this prior to submission of this patch, but 
looking at this again I think
you should use the get_update_locks_for_mdev() function for two reasons:

1. It is safer because it will take the matrix_dev->guests_lock which 
will prevent the matrix_mdev->kvm
     field from changing before you check it

2. I will eliminate the need for the else

get_update_locks_for_mdev(matrix_mdev)
if (matrix_mdev->kvm) {
     clear the masks
     signal guest config changed
}
...
release_update_locks_for_mdev(matrix_mdev); Sorry about not seeing this 
before you posted this patch.
> +		kvm_arch_crypto_clear_masks(matrix_mdev->kvm);
> +		signal_guest_ap_cfg_changed(matrix_mdev);
> +	} else {
> +		mutex_lock(&matrix_dev->mdevs_lock);
> +	}
> +
>   	if (matrix_mdev->req_trigger) {
>   		if (!(count % 10))
>   			dev_notice_ratelimited(dev,
> @@ -2068,6 +2075,12 @@ static void vfio_ap_mdev_request(struct vfio_device *vdev, unsigned int count)
>   		dev_notice(dev,
>   			   "No device request registered, blocked until released by user\n");
>   	}
> +
> +	if (matrix_mdev->kvm)
> +		release_update_locks_for_kvm(matrix_mdev->kvm);
> +	else
> +		mutex_unlock(&matrix_dev->mdevs_lock);
> +
>   }
>   
>   static int vfio_ap_mdev_get_device_info(unsigned long arg)

Re: [RFC PATCH v1] fixup! s390/vfio-ap: Notify userspace that guest's AP config changed when mdev removed

[Anthony Krowiak]

On 3/4/25 9:18 AM, Rorie Reyes wrote:
>
>
> On 3/4/25 8:36 AM, Anthony Krowiak wrote:
>>>     static void vfio_ap_mdev_update_guest_apcb(struct ap_matrix_mdev 
>>> *matrix_mdev)
>>> @@ -1870,7 +1870,6 @@ static void vfio_ap_mdev_unset_kvm(struct 
>>> ap_matrix_mdev *matrix_mdev)
>>>           get_update_locks_for_kvm(kvm);
>>>             kvm_arch_crypto_clear_masks(kvm);
>>> -        signal_guest_ap_cfg_changed(matrix_mdev);
>>>           vfio_ap_mdev_reset_queues(matrix_mdev);
>>>           kvm_put_kvm(kvm);
>>>           matrix_mdev->kvm = NULL;
>>> @@ -2057,6 +2056,14 @@ static void vfio_ap_mdev_request(struct 
>>> vfio_device *vdev, unsigned int count)
>>>         matrix_mdev = container_of(vdev, struct ap_matrix_mdev, vdev);
>>>   +    if (matrix_mdev->kvm) {
>>> +        get_update_locks_for_kvm(matrix_mdev->kvm);
>>
>> I know we talked about this prior to submission of this patch, but 
>> looking at this again I think
>> you should use the get_update_locks_for_mdev() function for two reasons:
>>
>> 1. It is safer because it will take the matrix_dev->guests_lock which 
>> will prevent the matrix_mdev->kvm
>>     field from changing before you check it
>>
> So I'll replace *get_update_locks_for_kvm(matrix_mdev->kvm)* with 
> *get_update_locks_for_mdev(&matrix_dev->guests_lock)*

See code below:get_update_locks_for_mdev(matrix_mdev)
That function will take the guests_lock

>> 2. I will eliminate the need for the else
>>
>> get_update_locks_for_mdev(matrix_mdev)
>> if (matrix_mdev->kvm) {
>>     clear the masks
>>     signal guest config changed
>> }
>> ...
>> release_update_locks_for_mdev(matrix_mdev); Sorry about not seeing 
>> this before you posted this patch.
>>> + kvm_arch_crypto_clear_masks(matrix_mdev->kvm);
>>> +        signal_guest_ap_cfg_changed(matrix_mdev);
>>> +    } else {
>>> +        mutex_lock(&matrix_dev->mdevs_lock);
>>> +    }
> So remove the else statement that contains the mutex function

Re: [RFC PATCH v1] fixup! s390/vfio-ap: Notify userspace that guest's AP config changed when mdev removed

[Rorie Reyes]

On 3/4/25 9:27 AM, Anthony Krowiak wrote:
>
>
>
> On 3/4/25 9:18 AM, Rorie Reyes wrote:
>>
>>
>> On 3/4/25 8:36 AM, Anthony Krowiak wrote:
>>>>     static void vfio_ap_mdev_update_guest_apcb(struct 
>>>> ap_matrix_mdev *matrix_mdev)
>>>> @@ -1870,7 +1870,6 @@ static void vfio_ap_mdev_unset_kvm(struct 
>>>> ap_matrix_mdev *matrix_mdev)
>>>>           get_update_locks_for_kvm(kvm);
>>>>             kvm_arch_crypto_clear_masks(kvm);
>>>> -        signal_guest_ap_cfg_changed(matrix_mdev);
>>>>           vfio_ap_mdev_reset_queues(matrix_mdev);
>>>>           kvm_put_kvm(kvm);
>>>>           matrix_mdev->kvm = NULL;
>>>> @@ -2057,6 +2056,14 @@ static void vfio_ap_mdev_request(struct 
>>>> vfio_device *vdev, unsigned int count)
>>>>         matrix_mdev = container_of(vdev, struct ap_matrix_mdev, vdev);
>>>>   +    if (matrix_mdev->kvm) {
>>>> +        get_update_locks_for_kvm(matrix_mdev->kvm);
>>>
>>> I know we talked about this prior to submission of this patch, but 
>>> looking at this again I think
>>> you should use the get_update_locks_for_mdev() function for two 
>>> reasons:
>>>
>>> 1. It is safer because it will take the matrix_dev->guests_lock 
>>> which will prevent the matrix_mdev->kvm
>>>     field from changing before you check it
>>>
>> So I'll replace *get_update_locks_for_kvm(matrix_mdev->kvm)* with 
>> *get_update_locks_for_mdev(&matrix_dev->guests_lock)*
>
> See code below:get_update_locks_for_mdev(matrix_mdev)
> That function will take the guests_lock
>
Ah ok, I see now. I'll make those changes. Thank you!
>>> 2. I will eliminate the need for the else
>>>
>>> get_update_locks_for_mdev(matrix_mdev)
>>> if (matrix_mdev->kvm) {
>>>     clear the masks
>>>     signal guest config changed
>>> }
>>> ...
>>> release_update_locks_for_mdev(matrix_mdev); Sorry about not seeing 
>>> this before you posted this patch.
>>>> + kvm_arch_crypto_clear_masks(matrix_mdev->kvm);
>>>> +        signal_guest_ap_cfg_changed(matrix_mdev);
>>>> +    } else {
>>>> +        mutex_lock(&matrix_dev->mdevs_lock);
>>>> +    }
>> So remove the else statement that contains the mutex function
>