Re: [PATCH v5 06/13] KVM: s390: interfaces to manage guest's AP matrix

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

On 05/16/2018 09:15 AM, Pierre Morel wrote:
> On 16/05/2018 15:12, Tony Krowiak wrote:
>> On 05/16/2018 03:48 AM, Pierre Morel wrote:
>>> On 15/05/2018 18:07, Tony Krowiak wrote:
>>>> On 05/15/2018 10:55 AM, Pierre Morel wrote:
>>>>> On 07/05/2018 17:11, Tony Krowiak wrote:
>>>>>> Provides interfaces to manage the AP adapters, usage domains
>>>>>> and control domains assigned to a KVM guest.
>>>>>>
>>>>>> The guest's SIE state description has a satellite structure 
>>>>>> called the
>>>>>> Crypto Control Block (CRYCB) containing three bitmask fields
>>>>>> identifying the adapters, queues (domains) and control domains
>>>>>> assigned to the KVM guest:
>>>>>>
>>> ...snip...
>>>>>> +}
>>>>>
>>>>> This function (ap_validate_queue_sharing) only verifies that VM 
>>>>> don't share queues.
>>>>> What about the queues used by a host application?
>>>>
>>>> How can that be verified from this function? I suppose I could put 
>>>> a check in here to
>>>> verify that the queues are reserved by the vfio_ap device driver, 
>>>> but that would
>>>> be redundant because an AP queue can not be assigned to a mediated 
>>>> matrix device
>>>> via its sysfs attributes unless it is reserved by the vfio_ap 
>>>> device driver (see
>>>> patches 7, 8 and 9).
>>>>
>>>>>
>>>>>
>>>>> I understand that you want to implement  these checks within KVM 
>>>>> but this is
>>>>> related to which queue devices are bound to the matrix and which 
>>>>> one are not.
>>>>
>>>> See my comments above and below about AP queue assignment to the 
>>>> mediated matrix
>>>> device. The one verification we can't do when the devices are 
>>>> assigned is whether
>>>> another guest is using the queue because assignment occurs before 
>>>> the guest using
>>>> the queue is started in which case we have no access to KVM. It 
>>>> makes no sense to
>>>> do so at assignment time anyway because it doesn't matter until the 
>>>> guest using
>>>> the mediated matrix device is started, so that check is done in KVM.
>>>>
>>>>>
>>>>>
>>>>> I think that this should be related somehow to the bounded queue 
>>>>> devices and
>>>>> therefor implemented inside the matrix driver.
>>>>
>>>> As I stated above, when an AP queue is assigned to the mediated 
>>>> matrix device via
>>>> its sysfs attributes, a check is done to verify that it is bound to 
>>>> the vfio_ap
>>>> device driver (see patches 7, 8 and 9). If not, then assignment 
>>>> will be rejected;
>>>> therefore, it will not be possible to configure a CRYCB with AP 
>>>> queues that are
>>>> not bound to the device driver.
>>>
>>> This patch and te followed patches take care that the queues are 
>>> bound to the
>>> matrix driver when they are assigned to the matrix using the sysfs 
>>> entries.
>>>
>>> But they do not take care that the queue can not be unbound before 
>>> you start
>>> the guest, and they are not in the path if the admin decide to 
>>> unbind a queue
>>> at some later time.
>>
>> That is a good point. I need to put a check in the device driver at 
>> the time
>> the mediated device fd is opened to verify that the queues being 
>> configured in
>> the guest's CRYCB are bound to the driver.
>
> not only, you also need to avoid the possibility of unbinding the device.
> For this you need to use the remove callback from the driver.

I thought I addressed this already. The definition of the remove 
callback does
not specify a return value, so there is currently no way to prevent the 
AP bus
from removing the queue device on unbind. I sent an email to Harald to 
discuss
adding a return value to the callback.

>
>
>>
>>>
>>>
>>>>
>>>>>
>>>>>
>>>>> Regards,
>>>>>
>>>>> Pierre
>>>>>
>>>>
>>>
>>
>