Re: [PATCH v3] s390/vfio-ap: clean up vfio_ap resources when KVM pointer invalidated

[Christian Borntraeger]

On 16.12.20 10:58, Christian Borntraeger wrote:
> On 16.12.20 02:21, Halil Pasic wrote:
>> On Tue, 15 Dec 2020 19:10:20 +0100
>> Christian Borntraeger <borntraeger@xxxxxxxxxx> wrote:
>>
>>>
>>>
>>> On 15.12.20 11:57, Halil Pasic wrote:
>>>> On Mon, 14 Dec 2020 11:56:17 -0500
>>>> Tony Krowiak <akrowiak@xxxxxxxxxxxxx> wrote:
>>>>
>>>>> The vfio_ap device driver registers a group notifier with VFIO when the
>>>>> file descriptor for a VFIO mediated device for a KVM guest is opened to
>>>>> receive notification that the KVM pointer is set (VFIO_GROUP_NOTIFY_SET_KVM
>>>>> event). When the KVM pointer is set, the vfio_ap driver takes the
>>>>> following actions:
>>>>> 1. Stashes the KVM pointer in the vfio_ap_mdev struct that holds the state
>>>>>    of the mediated device.
>>>>> 2. Calls the kvm_get_kvm() function to increment its reference counter.
>>>>> 3. Sets the function pointer to the function that handles interception of
>>>>>    the instruction that enables/disables interrupt processing.
>>>>> 4. Sets the masks in the KVM guest's CRYCB to pass AP resources through to
>>>>>    the guest.
>>>>>
>>>>> In order to avoid memory leaks, when the notifier is called to receive
>>>>> notification that the KVM pointer has been set to NULL, the vfio_ap device
>>>>> driver should reverse the actions taken when the KVM pointer was set.
>>>>>
>>>>> Fixes: 258287c994de ("s390: vfio-ap: implement mediated device open callback")
>>>>> Signed-off-by: Tony Krowiak <akrowiak@xxxxxxxxxxxxx>
>>>>> ---
>>>>>  drivers/s390/crypto/vfio_ap_ops.c | 29 ++++++++++++++++++++---------
>>>>>  1 file changed, 20 insertions(+), 9 deletions(-)
>>>>>
>>>>> diff --git a/drivers/s390/crypto/vfio_ap_ops.c b/drivers/s390/crypto/vfio_ap_ops.c
>>>>> index e0bde8518745..cd22e85588e1 100644
>>>>> --- a/drivers/s390/crypto/vfio_ap_ops.c
>>>>> +++ b/drivers/s390/crypto/vfio_ap_ops.c
>>>>> @@ -1037,8 +1037,6 @@ static int vfio_ap_mdev_set_kvm(struct ap_matrix_mdev *matrix_mdev,
>>>>>  {
>>>>>  	struct ap_matrix_mdev *m;
>>>>>
>>>>> -	mutex_lock(&matrix_dev->lock);
>>>>> -
>>>>>  	list_for_each_entry(m, &matrix_dev->mdev_list, node) {
>>>>>  		if ((m != matrix_mdev) && (m->kvm == kvm)) {
>>>>>  			mutex_unlock(&matrix_dev->lock);
>>>>> @@ -1049,7 +1047,6 @@ static int vfio_ap_mdev_set_kvm(struct ap_matrix_mdev *matrix_mdev,
>>>>>  	matrix_mdev->kvm = kvm;
>>>>>  	kvm_get_kvm(kvm);
>>>>>  	kvm->arch.crypto.pqap_hook = &matrix_mdev->pqap_hook;
>>>>> -	mutex_unlock(&matrix_dev->lock);
>>>>>
>>>>>  	return 0;
>>>>>  }
>>>>> @@ -1083,35 +1080,49 @@ static int vfio_ap_mdev_iommu_notifier(struct notifier_block *nb,
>>>>>  	return NOTIFY_DONE;
>>>>>  }
>>>>>
>>>>> +static void "(struct ap_matrix_mdev *matrix_mdev)
>>>>> +{
>>>>> +	kvm_arch_crypto_clear_masks(matrix_mdev->kvm);
>>>>> +	matrix_mdev->kvm->arch.crypto.pqap_hook = NULL;
>>>>
>>>>
>>>> This patch LGTM. The only concern I have with it is whether a
>>>> different cpu is guaranteed to observe the above assignment as
>>>> an atomic operation. I think we didn't finish this discussion
>>>> at v1, or did we?
>>>
>>> You mean just this assigment:
>>>>> +	matrix_mdev->kvm->arch.crypto.pqap_hook = NULL;
>>> should either have the old or the new value, but not halve zero halve old?
>>>
>>
>> Yes that is the assignment I was referring to. Old value will work as well because
>> kvm holds a reference to this module while in the pqap_hook.
>>  
>>> Normally this should be ok (and I would consider this a compiler bug if
>>> this is split into 2 32 bit zeroes) But if you really want to be sure then we
>>> can use WRITE_ONCE.
>>
>> Just my curiosity: what would make this a bug? Is it the s390 elf ABI,
>> or some gcc feature, or even the C standard? Also how exactly would
>> WRITE_ONCE, also access via volatile help in this particular situation?
> 
> I think its a tricky things and not strictly guaranteed, but there is a lot
> of code that relies on the atomicity of word sizes. see for example the discussion
> here
> https://lore.kernel.org/lkml/CAHk-=wgC4+kV9AiLokw7cPP429rKCU+vjA8cWAfyOjC3MtqC4A@xxxxxxxxxxxxxx/
> 
> WRITE_ONCE will not change the guarantees a lot, but it is mostly a documentation
> that we assume atomic access here.

After looking again at the code, I think I have to correct myself.
WRITE_ONCE does not look necessary.


Another thing, though:
Shouldnt we also replace this code

[...]
static void vfio_ap_mdev_release(struct mdev_device *mdev)
{
        struct ap_matrix_mdev *matrix_mdev = mdev_get_drvdata(mdev);

        mutex_lock(&matrix_dev->lock);
        if (matrix_mdev->kvm) {
--->          kvm_arch_crypto_clear_masks(matrix_mdev->kvm);
--->          matrix_mdev->kvm->arch.crypto.pqap_hook = NULL;
--->          vfio_ap_mdev_reset_queues(mdev);
--->          kvm_put_kvm(matrix_mdev->kvm);
--->          matrix_mdev->kvm = NULL;
[...]

with vfio_ap_mdev_unset_kvm ?