Date:   Tue, 17 Apr 2018 12:08:53 -0400
From:   Tony Krowiak <akrowiak@...ux.vnet.ibm.com>
To:     Pierre Morel <pmorel@...ux.vnet.ibm.com>,
        linux-s390@...r.kernel.org, linux-kernel@...r.kernel.org,
        kvm@...r.kernel.org
Cc:     freude@...ibm.com, schwidefsky@...ibm.com,
        heiko.carstens@...ibm.com, borntraeger@...ibm.com,
        cohuck@...hat.com, kwankhede@...dia.com,
        bjsdjshi@...ux.vnet.ibm.com, pbonzini@...hat.com,
        alex.williamson@...hat.com, alifm@...ux.vnet.ibm.com,
        mjrosato@...ux.vnet.ibm.com, jjherne@...ux.vnet.ibm.com,
        thuth@...hat.com, pasic@...ux.vnet.ibm.com, berrange@...hat.com,
        fiuczy@...ux.vnet.ibm.com, buendgen@...ibm.com
Subject: Re: [PATCH v4 13/15] KVM: s390: configure the guest's AP devices

On 04/16/2018 09:05 AM, Pierre Morel wrote:
> On 15/04/2018 23:22, Tony Krowiak wrote:
>> Registers a group notifier during the open of the mediated
>> matrix device to get information on KVM presence through the
>> VFIO_GROUP_NOTIFY_SET_KVM event. When notified, the pointer
>> to the kvm structure is saved inside the mediated matrix
>> device. Once the VFIO AP device driver has access to KVM,
>> access to the APs can be configured for the guest.
>>
>> Access to APs is configured when the file descriptor for the
>> mediated matrix device is opened by userspace. The items to be
>> configured are:
>>
>> 1. The ECA.28 bit in the SIE state description determines whether
>>     AP instructions are interpreted by the hardware or intercepted.
>>     The VFIO AP device driver relies interpretive execution of
>>     AP instructions so the ECA.28 bit will be set
>>
>> 2. Guest access to AP adapters, usage domains and control domains
>>     is controlled by three bit masks referenced from the
>>     Crypto Control Block (CRYCB) referenced from the guest's SIE state
>>     description:
>>
>>     * The AP Mask (APM) controls access to the AP adapters. Each bit
>>       in the APM represents an adapter number - from most significant
>>       to least significant bit - from 0 to 255. The bits in the APM
>>       are set according to the adapter numbers assigned to the mediated
>>       matrix device via its 'assign_adapter' sysfs attribute file.
>>
>>     * The AP Queue (AQM) controls access to the AP queues. Each bit
>>       in the AQM represents an AP queue index - from most significant
>>       to least significant bit - from 0 to 255. A queue index references
>>       a specific domain and is synonymous with the domian number. The
>>       bits in the AQM are set according to the domain numbers assigned
>>       to the mediated matrix device via its 'assign_domain' sysfs
>>       attribute file.
>>
>>     * The AP Domain Mask (ADM) controls access to the AP control 
>> domains.
>>       Each bit in the ADM represents a control domain - from most
>>       significant to least significant bit - from 0-255. The
>>       bits in the ADM are set according to the domain numbers assigned
>>       to the mediated matrix device via its 'assign_control_domain'
>>       sysfs attribute file.
>>
>> Signed-off-by: Tony Krowiak <akrowiak@...ux.vnet.ibm.com>
>> ---
>>   drivers/s390/crypto/vfio_ap_ops.c     |   50 
>> +++++++++++++++++++++++++++++++++
>>   drivers/s390/crypto/vfio_ap_private.h |    2 +
>>   2 files changed, 52 insertions(+), 0 deletions(-)
>>
>> diff --git a/drivers/s390/crypto/vfio_ap_ops.c 
>> b/drivers/s390/crypto/vfio_ap_ops.c
>> index bc2b05e..e3ff5ab 100644
>> --- a/drivers/s390/crypto/vfio_ap_ops.c
>> +++ b/drivers/s390/crypto/vfio_ap_ops.c
>> @@ -53,6 +53,54 @@ static int vfio_ap_mdev_remove(struct mdev_device 
>> *mdev)
>>       return 0;
>>   }
>>
>> +static int vfio_ap_mdev_group_notifier(struct notifier_block *nb,
>> +                       unsigned long action, void *data)
>> +{
>> +    struct ap_matrix_mdev *matrix_mdev;
>> +
>> +    if (action == VFIO_GROUP_NOTIFY_SET_KVM) {
>> +        matrix_mdev = container_of(nb, struct ap_matrix_mdev,
>> +                       group_notifier);
>> +        matrix_mdev->kvm = data;
>> +    }
>> +
>> +    return NOTIFY_OK;
>> +}
>> +
>> +static int vfio_ap_mdev_open(struct mdev_device *mdev)
>> +{
>> +    struct ap_matrix_mdev *matrix_mdev = mdev_get_drvdata(mdev);
>> +    unsigned long events;
>> +    int ret;
>> +
>> +    matrix_mdev->group_notifier.notifier_call = 
>> vfio_ap_mdev_group_notifier;
>> +    events = VFIO_GROUP_NOTIFY_SET_KVM;
>> +
>> +    ret = vfio_register_notifier(mdev_dev(mdev), VFIO_GROUP_NOTIFY,
>> +                     &events, &matrix_mdev->group_notifier);
>> +    if (ret)
>> +        return ret;
>> +
>> +    ret = kvm_ap_interpret_instructions(matrix_mdev->kvm, true);
>
> Do you need this call ?
> apie is always enabled in KVM if AP instructions are available.

I suppose we don't, in which case we don't need the 
kvm_ap_interpret_instructions()
function either ... at least not until we implement interception.

>
>
> Setting or not the interpretation is done for the VM in a all.
> It is not the right place to do it here since open is device dependent.

As I stated above, at this time we probably do not need this, however;
that will not always be the case. The setting is and always will be for the
VM in all - unless the architecture changes - because it is controlled by a
single bit (ECA.28). If you recall, I originally set interpretation in the
vfio_ap device driver when notified of the VFIO_GROUP_NOTIFY_SET_KVM event.
I believe ultimately that it is the device driver that should set the value
for apie.



>
>
> Or we only have one device in the VM at a time.
> In this case, shouldn't we make it official by returning -EEXIST for 
> the second call?

We do allow only one vfio-ap device at a time. QEMU will allow only one 
vfio-ap device
to be configured for a guest. Should we also put a check in here?

>
>
>
>> +    if (ret)
>> +        return ret;
>> +
>> +    ret = kvm_ap_configure_matrix(matrix_mdev->kvm,
>> +                      matrix_mdev->matrix);
>> +
>> +    return ret;
>> +}
>> +
>> +static void vfio_ap_mdev_release(struct mdev_device *mdev)
>> +{
>> +    struct ap_matrix_mdev *matrix_mdev = mdev_get_drvdata(mdev);
>> +
>> +    kvm_ap_deconfigure_matrix(matrix_mdev->kvm);
>> +    kvm_ap_interpret_instructions(matrix_mdev->kvm, false);
>
> This call clears the apie in KVM.
> This is only OK if we have a single device present until the end of 
> the VM,
> otherwise AP instructions in the guest will fail after the release 
> until the end of the VM
> or until a new device is plugged.

See Message ID: 
<1523819244-29954-5-git-send-email-akrowiak@...ux.vnet.ibm.com> on the
qemu mailing list. There will be only one vfio-ap device allowed for the 
MVP model.

>
>
>
>> + vfio_unregister_notifier(mdev_dev(mdev), VFIO_GROUP_NOTIFY,
>> +                 &matrix_mdev->group_notifier);
>> +}
>> +
>>   static ssize_t name_show(struct kobject *kobj, struct device *dev, 
>> char *buf)
>>   {
>>       return sprintf(buf, "%s\n", VFIO_AP_MDEV_NAME_HWVIRT);
>> @@ -754,6 +802,8 @@ static ssize_t matrix_show(struct device *dev, 
>> struct device_attribute *attr,
>>       .mdev_attr_groups    = vfio_ap_mdev_attr_groups,
>>       .create            = vfio_ap_mdev_create,
>>       .remove            = vfio_ap_mdev_remove,
>> +    .open            = vfio_ap_mdev_open,
>> +    .release        = vfio_ap_mdev_release,
>>   };
>>
>>   int vfio_ap_mdev_register(struct ap_matrix *ap_matrix)
>> diff --git a/drivers/s390/crypto/vfio_ap_private.h 
>> b/drivers/s390/crypto/vfio_ap_private.h
>> index f248faf..48e2806 100644
>> --- a/drivers/s390/crypto/vfio_ap_private.h
>> +++ b/drivers/s390/crypto/vfio_ap_private.h
>> @@ -31,6 +31,8 @@ struct ap_matrix {
>>
>>   struct ap_matrix_mdev {
>>       struct kvm_ap_matrix *matrix;
>> +    struct notifier_block group_notifier;
>> +    struct kvm *kvm;
>>   };
>>
>>   static inline struct ap_matrix *to_ap_matrix(struct device *dev)
>
>

Powered by blists - more mailing lists